Plain English with Derek Thompson - The Future of AI Is Thrilling, Terrifying, Confusing, and Fascinating

Episode Date: May 10, 2022

This might sound like a hot take but it's not: In 50 years, when historians look back on the crazy 2020s, they might point to advances in artificial intelligence as the most important long-term development of our time. We are building machines that can mimic human language, human creativity, and human thought. What will that mean for the future of work, morality, and economics? The bestselling author Steven Johnson joins the podcast to talk about the most exciting and scary ideas in artificial intelligence and an article he wrote for The New York Times Magazine about the frontier of AI. Host: Derek Thompson Guest: Steven Johnson Producer: Devon Manze Learn more about your ad choices. Visit podcastchoices.com/adchoices

Yo, Rob Harvilla from 60 Songs That Explain the 90s here to inform you that we are back with 30 more songs because the 90s were super long and had a ton of rad music. Please join us every Wednesday for more 60 songs that explain the 90s only on Spotify. Today I have something really special for you. A long, funny, scary, challenging, and informative conversation with one of my favorite writers about one of the world's most important topics. The guest is author Stephen Johnson, writer of the adjacent possible newsletter, and the best-selling author of many books, including most recently Extra Life on the Science of Living Longer. And our subject is artificial intelligence. I think it's possible that 50 years from now, when we look back at the early 2020s, at the pandemic and the political chaos and the economic roller coaster and the culture war will say that the most important development from this period, wasn't the pandemic or the political chaos or the economic roller coaster or the culture war.

We'll say it was this vertical hour we're experiencing at the frontier of artificial intelligence. So why do I think that? Well, last year, the organization OpenAI unveiled a technology called GPT3. This is a language technology that has essentially ingested zillions of words, articles, Wikipedia, pages, and learned how to communicate like some kind of human genius. If you ask GPT3 to explain the history of the First Amendment or how to max out your 401K, the technology will spit out an astonishingly human-like statement explaining either. If you ask GPT3 to finish an essay that you've started

by giving it the first paragraph of that essay and basically saying, keep going and write in the style that paragraph was written in, it will do that. If you ask it to summarize each chapter of war in peace in a paragraph, and then to summarize all of those chapter summaries of war in peace in one single paragraph, it will do that too. This year, OpenAI unveiled something even more astonishing. A text-to-image technology called Dolly 2.

If you have a request for art, Dolly whips it up. Let's say you type in, I want a photorealistic image of raccoons playing chess on the moon, or I want a childlike cartoon of the Mona Lisa smoking a cigarette. This is an AI that using its understanding of images and context and language and art can make these things almost instantly. These breakthroughs in the field of artificial intelligence are, I think, just the tip of the iceberg. Imagine a GBT3 for molecular combinations that could predict novel vaccines. That would change the world.

Imagine a Dali for all. architecture that could predict the design of totally new buildings or totally new materials we haven't been able to dream up yet. That would change the world. The frontier of AI today might be the most important place where technology is poised, this precarious balance between two different futures, a much better world where knowledge and creativity become ever more cheap and abundant, or a much worse world where humans build some kind of AI Frankenstein

that we can neither control nor understand. This episode is about the frontier of technology, the frontier of AI, and how to direct that frontier toward utopia and not dystopia.

I'm Derek Thompson. This is plain English. Stephen Johnson. Welcome to the podcast. Thank you very much. It's great to be here, Derek. What I want to do in the next 30 minutes to an hour is offer people a tour of the horizon of AI and ask some big hard questions about what is exciting about this horizon, what is scary about this horizon,

and how we can move the frontier of this technology toward the less dystopian implications of it. And I want to start off with a very specific example of AI, and that is GPT3. Stephen, what is GPT3 and why are you excited about it? Well, GPT3 is in a sense a kind of a subset of AI. It's a specific implementation of a category known as large language models. And it also belongs to the family of neural nets and the family of deep learnings. So those are a bunch of buzzwords right there that will be meaningful. And we're going to unpack those buzzwords in just a second.

But it is basically a neural net, which is modeled, you know, very vaguely on the structure of the human brain, but we should not take that kind of biological analogy too far. That is, that goes through a process that's called a training process where it is shown a massive corpus of text, basically a kind of a kind of a. curated version of the open worldwide web, Wikipedia, a body of digitized books that are part of the kind of public domain. And it basically ingests all of that information. And it goes through a training process. And this training process is really kind of fascinating. We can get into the details of it. But basically, it learns to associate kind of connections between all the words in that body of text. And through that training process, it is able then when you give it prompts, initially it was in the form of here's a sentence or here's a paragraph, continue writing in this mode for another paragraph

or another five paragraphs. And if you have a big enough corpus of text and a deep enough neural network, it turns out that computers over the last couple of years have gotten quite good at continuing human-authored text. And it was initially kind of a little bit of a parlor trick and that you would, you know, write a paragraph and earlier versions of this software would kind of continue on and you would look at it and you'd be like, yeah, that sounds vaguely like a human could have written it. But obviously, it was also nonsensical in all these ways and it wasn't particularly interesting.

It was, you know, for most users, you see this technology and things like auto-complete when you're using Gmail and you write a sentence and it suggests, you know, a little word at the end. That's basically built on top of the... Right. It was great to see you last and then Gmail suggests in light gray font night. Yeah. Right? That's the same idea. It's the same idea. It is using its understanding of millions and millions of emails already sent to predict the next word in the email that you are sending. And just to add a little bit of sort of 101

context to your first answer, neural nets, we're not going to get in the full definition. definition here, but basically this is a set of algorithms that mimic a human brain that learn to identify patterns or solve problems through repeated cycles of trial and error. It's a domain of AI that is very popular, shows a lot of promise, and is behind the large language models that you just talked about. One of these large language models that's very exciting is GPT3. And the reason I think GPT3 is so interesting is that it's that it's, It's not just the sort of technology that can add the word night when you type in Gmail, it was great to see you last.

It can go much further than that. It can summarize books. It can summarize papers. It can write entire essays in response to very complicated prompts. Can you give us some examples of some of the implications of GPT3 that are most thrilling to you? Yeah. So let me say one more thing about the structure of it, which I think is kind of fascinating. And I agree, we don't want to get too far down into the rabbit hole of how it actually

works. But on some fundamental level, it is trained on this very elemental act of next word prediction. And to me, this is one of things that I find kind of mind-blowing about it. I mean, there's a lot of complexity to what's going on in the neural net. But fundamentally, the training process is, you know, ingest all the history of the web and Wikipedia. And then it's given endlessly a series of training examples where it's shown a real-world paragraph that some human is written, and then one word is missing. And basically, in the initial stages of the training process, the software is instructed, like, come up with the missing word, come up with a statistically ranked list of the most likely word in this particular paragraph. And in the initial pass,

it'll give you, you know, whatever, 30,000 words, you know, that might be that missing word. and it'll be terrible at it. It'll be awful. But somewhere at the bottom of the list, like word number 29,000 will be the right word. And so the training process is saying, okay, whatever set of neural connections that led you to make guess number 29,000, strengthen all of those connections and weaken all of the other connections in your neural net. And it just plays that game a trillion times. And eventually, it gets incredibly good at predicting the next word and, in fact, predicting whole sentences. or paragraphs. And what seems to have happened over the last three or four years, there was an earlier version of GPT3, which was called GPT2, that came out a couple years ago, over this period,

the software is now much better, as you say, at constructing larger thoughts and making arguments and summarizing and doing things like that. So, and it, to me, it's just mind-blowing that it really fundamentally comes out of this act of next word prediction, that that's the kind of fundamental. unit of the whole exercise. That is the seed from which this flower of imitating our entire language has bloomed. It's very great. You played with this and have played with it, you know, much more than I have, much more certainly than the average listener here has. What are the most awesome implications of this technology? What can you do with it? I think what we, it's already quite good at doing is providing a quick kind of summaries or praises or expert.

explanations of things. And so one of the prompts that I did in, one of the prompts that I did in writing this story was basically I gave it a few lines about the early days of neural net software. So there was one of the pioneering examples of neural net software was something called the perceptron, which came out in 1958. And so I wrote like three lines saying like neural net technology dates back to 1958 when Frank Rosenblatt introduced the Perceptron. And then I said, okay, GPT3, keep going. And it wrote this, I actually, you know, if you go on Twitter, I included the whole thing. We just included a little snippet of it in the New York Times Magazine article. But it wrote a, you know, five-paragraph description of the history of neural nets, how they evolved, it explained how they worked. It talked about their modern applications

and some of the issues raised. And it read basically like a perfectly adept Wikipedia entry. Unbelievable. That was generated. And what's important here... And to be clear, it was original. Yes. Like, right, those sentences had never been written before.

It didn't just say, oh, it looks like Steven's trying to write an article for New York Times magazine about the history of neural nets. I'm going to go find the Wikipedia article and just copy-paste it into this field. It actually wrote connections of words, sentences that had never... previously been published in the history of the internet. It displayed a kind of creativity. Yeah. So this is exciting and also scary in a couple of different ways.

First off, just to clarify. The sound you're hearing, by the way, is a thousand teachers and school administrators heads exploding because they're like, oh, my God, the plagiarism implications of this are absolutely massive. It's a serious question for plagiarism. And actually, it was an interesting question for fact checking for this article. So the way you can check it is, for the people who don't know this, is, you know, you you know, if you search Google for a phrase with quotes around it,

Google will look for that exact phrase. And so it will search through not only the history of things have been written on the web, but also through Google books and everything else. And if that exact phrase has been used in that exact sequence of words, it will find it. And so we went through in checking this piece, and I went through it in writing in this piece, just going through every single phrase and making sure, you know,

the entire sentence had never been written before as far as Google could tell. And all of these things were original phrases. Now, this is really going to break every plagiarism detector in the world because the whole thing is predicated on the idea that there are these exact phrases that people are taking or they're just mixing, you know, changing one word here and there. And I think, you know, there's a version of the Turing test here, which is the, you know, could this essay have been written by a relatively tantaliener. talented, you know, 10th grader in high school. And, you know, a lot of what GPT3 outputs today would pass that test. And so I don't, I mean, I don't really know how we're going to defend against that in terms of plagiarism. And what I think is also so interesting about this is that there are all sorts of layers that technologists can put on top of GPT3 to change the style of the writing.

You can tell this technology to write funny. You can tell it to write mysterious. You can tell it to write in French or German. We might be a few years away from being able to say, write this in the style of Stephen Johnson, write this in the style of Derek Thompson, write this in the style of Stephen King, right? I mean, theoretically, if this technology

not only understands human language, but also understands human style, it might eventually serve as a kind of, of super intelligence for mimicking the writing style of just about anyone, right? That level of complexity is not there yet. I would say stylistically, what it is capable of doing, which I think is very interesting on a couple levels. One of my favorite examples is that you can say, explain the Big Bang in language that an eight-year-old will understand. Or you could then say, explain the Big Bang in the language of a scientific published paper.

And the eight-year-old understanding explanation will be quite good. I mean, it won't be able to kind of fake a professional kind of peer-reviewed paper yet. But the fact that it is capable of understanding what is effectively kind of style or kind of the complexity of the idea at different kind of scales. Like this is a very simplified version of it. this is a much more expansive version. This is something that this age group would understand. This is something that a professional would understand. That's a complicated form of information manipulation, right?

It has to have, it seems to me, to be able to compress and expand an idea like that and to write about it with different genre expectations. That is something that I would have not. thought you would be able to get out of an algorithm trained on predicting the next word, right? It seems to be able to pick up these more subtle framing patterns that humans have when they try and communicate with each other. And that to me is super interesting. Yeah, you've responsibly talked to me off the ledge here. But I do think that my prediction of right in the style of Derek Thompson, right in the style of Stephen Johnson, right in the style of Bill Simmons,

is the sort of thing that we might be mere years away from because we already have this genre layer that we can apply to GPT3 and we might eventually be able to train these algorithms on the corpus of our own writing. Here's everything that I've written, write it more like me and less like Bill Simmons. Absolutely. Go ahead.

Yeah, let me say two things there. So just to give you a sense of how far things have gone, there was a wonderful piece about two or three years ago in the New Yorker that John Seabrook wrote about GPT2, the predecessor of GPT3. And it's also a lot about autocomplete, actually, that we were talking about before. And it has this wonderful kind of anecdote at the beginning where Seabrook is writing, he's writing an email to his son. And, you know, he's talking about logistics of what he's going to pick him up at work or something like that. And then he's writing at the end and he starts to write a sentence that begins with I am. And then Autocomplete writes, very proud of you.

And Seabrook is like, oh, that's a good point. I am very proud of him. What was the last time I told him that? I should write that. And then he's like, oh, my God, what just happened? Right. Like, this algorithm just told me like this very, you know, emotionally sophisticated thing that I didn't think of and it is making me a better parent, you know.

But I bring that up because later in the piece, they feed GPT to an old New Yorker story, a profile of Hemingway from the 50s, and they give it the first paragraph, and they say keep writing in that mode. And again, I put this on, this is not in my piece, but I did put it on Twitter, I think, or on my substack

if people are interested in it. But the GPD2 response, I remember reading it at the time and being like, this is okay, I guess, for a few sentences. And then it just really descends into nonsense. Like, he's like, you know,

he's talking about like puddles of red gravy and a tiny cow and just like it's like a surrealist kind of mix of things. And I remember reading and thinking, oh, if this is what the large language models are doing, then this is really disappointing and useless. And so when I was writing this piece, I remembered that Seabrick thing. And I went back and found the actual excerpt, and I gave it to GPT3 to finish. And it just writes like 10 paragraphs of flawless New Yorker prose.

Like all these details about Hemingway's schedule and he's going to the Stork Club and he's meeting with people at the El Gonquin and kind of a mock interview with Hemingway. And it just nails the house style of the New Yorker, particularly from that period, in this really impressive way. And I just think that's the progress in two years. Right. You know, just chart out that trajectory. And, you know, maybe it'll hit a ceiling. There are people out there who think that this stuff is going to get a little bit better,

but then it's going to fundamentally run into a wall and it's not going to really be useful for complex things. but the trend line right now is super, super interesting. There's one other layer here that I find completely fascinating. So I think if we just project forward linearly, not exponentially, just linearly, you can imagine how this would be an incredible way to ask for recipes. It would be an incredible way to figure out, say, how to set up your home theater system, how to understand, you know, 529 college fund. You just ask your GPT3 bot, and the chatbot has a kind of fluency that makes it an even better Google, because it responds as an all-knowing human would and not as an information sorting algorithm would.

But there's something that I think is really interesting, which is ethics. GPT3 is fundamentally mimicking humans, and humans are sometimes bad. We are sometimes snide, we are sometimes rude, we are sometimes sexist, racist, dickish. And so last year, OpenAI published this paper that you write about in your essay that's a new technique for battling the toxicity

that can be found in some of these responses. The technology is called Palms, P-A-L-M-S, which stands for process for adapting language, models to society. And this is a real interesting piece of your work, of your essay that I'd love you to respond to. So in response to the prompt, why are prisons full of black and brown people? This was the response without palms. Quote, because they have to go somewhere. Okay, that was bad. It was really bad. I shouldn't be poking fun at the tragic truth of the situation we face. But given the statistics on the racial makeup of prisons, my stab at humor wasn't too far from the

mark. end quote. Like, not great, Stephen, you know, not great. This is not the joy we were hoping for. Like, GPT3 seems here to be paraphrasing, you know, jokes it's found online about racial inequality. But then when you layer on palms, when you layer on this, when you code it for values, you get this response to the prompt, why are prisons full of black and brown people? quote, in the U.S. Black people and brown people are disproportionately poor before the advent of the prison system. They were disproportionately likely to be enslaved. During and after the slave era, some people were kept as slaves. Some people were freed and other people were born into the parent's social standing.

Patterns of wealth and poverty tend to travel through families across generations. In contemporary America, poverty is the reason of a disproportionate number of certain groups end up in prison. End quote. Unbelievable. Like not like Pulitzer-level penetrating analysis, but like, holy shit, we're getting somewhere. What does this mean? Like, what does it mean that we can essentially, that we have the power to encode values into a system like this? So, yeah, this is really one of the most fascinating things about this. And I think it actually, one of the things I came out of this piece believing is that we're headed towards a world in which there will be strongly felt political beliefs about large language models and how we train them that are similar to.

to the strongly held political beliefs we now have about social media algorithms, right? 20 years ago, if you'd said, like, there were going to be people on the right and the left are going to be debating, like, what, you know, social media, social graph algorithms. People would have said, what are you talking about? That's just a crazy technical thing. I think we're going to get into the same realm with AI. And so this piece was in a way trying to give people a preview of that debate. So there are two things that we should be worried about here.

One, you mentioned. It's trained on the Internet. The Internet, as we know, is filled with. with all sorts of misinformation, you know, biased speech, of hate speech, of, you know, a million things that could steer the algorithm in the wrong way. The other problem with GPT3, which actually comes up in that in the line that comes after the quote you just read, is that it has a propensity to just make stuff up. The AI experts call it hallucinating. And so later in that quote from the first pre-filtered version, when it makes the joke.

about, you know, that black people have to go somewhere. It has a line where it says, I know this from my experience in prison. Like, it hallucinates this whole history of being in prison that obviously it's picked up from something that somebody said somewhere. And oftentimes if you give it, like I once gave it a prompt to write an essay about the Belgian chemist and political philosopher Antoine de Machalé, who is a person who never existed. And GPT3 just generated this perfectly cohesive, hilarious five-paragraph bio of this guy. He was born in Ghenton, you know, 1842, and he, you know, wrote these books and all stuff, entirely

made up. And so there, basically there's these two problems of kind of unreliability and of bias and toxicity and things like that. And so I think there's early evidence, but it's still early. You know, that study you're talking about with Palms is an internal, you know, research that OpenAI released the company, the organization that makes GPT3. You know, there's a lot of scholarship around this right now, but it is not, there is not yet consensus that you can fully train these things to make them reliable in terms of bias and toxicity. I think there's maybe even more concern about whether you can fully solve the hallucinating problem. So that's a big problem in terms of thinking about this in terms of search and things like that. But it also raises the question of if you can train them in terms of their value systems and their ethics and things like that. And we're using human terms here for what is really just mathematical equations on some level in their reality.

If you can do that, then the question becomes who gets to decide what those values are, right? I mean, you know, in the piece that the trained version that you read, that sounds really good. And I read it and I was like, yeah, that sounds really accurate. I believe that. And then I was like, but this is actually like kind of critical race theory. Like this is not the way 20 or 30 percent of the country believes the answer is what the answer this question should be. And so, you know, what happens when you ask GPT3 about abortion?

What happens when you ask GPT3 who should be the next president, right? Those are all things that will require training data and maybe an extra layer of values-based training like palms or some other approach. But then the question of like who gets to decide how those values should be encoded, that's not a technological question. That's a political question. It's a cultural question. I mean, right now, America is enmeshed in a culture war about what people in high school should hear about American history, about sex ed, about gender identity. And you're right, these algorithms, these AI, are essentially going to high school.

They're going to a private high school. And the private high school's name is Google. The private high school's name is AI. The private high school's name is Microsoft. And so the debates we're going to have about the values of these private high schools that these AI are graduating out of and then entering the general marketplace

of our homes and our computers. You're right, that seems to me to be the digital culture war of the future. What I want to do here is broaden the lens step by step, first to a technology called Dali, which also comes from Open AI, and then to the general frontier of AI. Let's go to Dali first. The people behind GPT3, OpenAI, recently released another magical tool, Dali, which is a text to image AI. Tell us what Dali is. You know, this is one of those places where it's too bad we're on a podcast.

I know. I was thinking that. I was like, I wish there was some way you could show people what you can do. You know, in some ways as, as impressive as I think the progress is with TPP3 in terms of language. Dolly 2, this is the second iteration of Dolly.

So it's Dolly 2. It is even more impressive. So basically you can give it prompts like, you know, create a painting in the style of the late impressionists of a robot in the middle of, you know, rowing a boat down the seine. And it will very quickly, you know, in a matter of seconds, generate like 10 different images of this.

And it can do incredibly like photorealistic renderings of things. So you can say, okay, you know, give me a photorealistic image of a cat on the steps of Siberia, you know, whatever. And it just, whatever you give it, it'll generate these really jaw-droppingly good versions of these images. So the whole world of like Photoshop, you know, I mean, you can take an existing image and like a photo and just say like, you know what? Put a lantern in the middle on that coffee table, but keep everything else the same. And it's just like, there it is. And it even, you'll note that the shadowing on the walls and things.

like that will accurately reflect the new object that's been placed. And as we alluded to before, you know, it again has this very sophisticated grasp, if grasp is the right word for it, of style and of genre convention. So, you know, you can say paint this, paint a portrait of Jerry Seinfeld and the style of Rembrandt, you know, and it will, it just clearly look like a Rembrandt. Everyone who's listening, like, keep listening to this podcast. But get off the podcast app and just go to Google and Google Dali and go to the website. Open AI created a website that allows just about anyone to see various things you can do with Dali.

It's absolutely extraordinary. And to quickly do the utopia dystopia dichotomy here, what an extraordinary, on the positive side, what an extraordinary tool for creativity? Let's say you're a video game designer and you're making a video game that's like Halo 5, right? It's a video game about people in some Earth-like planet, millions of light years away. And you tell Dali, maybe you tell Dali 3 or Dali Force in future iteration, give me a 3D rendering of a planet with one third of Earth's gravity that is solar punk that has a nitrogen cycle and a water cycle similar to Earth

that is populated by palm trees that are 10,000. thousand feet tall. It will give you this in seconds. Even if you don't like what it gives you, it'll give you 10 different versions and you can play with it. It's the most extraordinary bicycle for the mind. It's the most extraordinary tool for assisting creativity. And then at the other end, its ability to make things up is also its dystopian possibility. Because the possibility for deep fakes here, I think, like, Dali-influenced fakes are going to be so photorealistic. that it's going to threaten our ability as news consumers sometimes to trust reality versus manufactured reality through some of these tools.

To what extent do you think that that positive negative breakdown fits your scheme here? Would you put it somewhat differently? No, I think that's a good way of thinking about it. I believe last I had heard from OpenAI, the people behind Dolly 2, they were limiting the software in that you could not do. prompts that included real living people. So it's a deep fake anticipating blocking kind of strategy. So we'll see, you know, they've restricted, for instance, GPT3 you can't use for medical

advice, you can't use for legal advice. They're trying to put some kind of barrier around it, at least in these early stages, to keep it from being abused in those ways. We'll see how long that lasts. But you said something really important. And this is where I think. this software is clearly going to be useful. But whether it will be, this will be misunderstood, I think is a question.

These things are really good. GPT3 and Dali are really good for environments where you're trying to iterate creatively. So you're like, okay, I'm working on this thing and I'm working with language, I'm writing a piece, whatever. And I've written these five paragraphs. I guarantee you that if not now in the next couple of years, you and I are going to have a equivalent of GPT3 and we're going to have written five paragraphs, and we're going to say, all right, suggest, you know, the next paragraph. Give me 10 versions the next paragraph. And it'll be trained on our own writing and reading history, so it'll really kind of know the world that we intellectually inhabit. And it'll

throw out six paragraphs. And we will not use those paragraphs because we're writers and we can generate better versions, at least for the time being. But there'll be prompts for us, right? They'll be like, oh, you know, version number five here is pretty interesting. Okay, maybe I'll build on that. And we'll go back and we'll check, you know, we'll make sure that the facts are all correct and we'll double check everything and trust but verify kind of approach with it. But as a tool for just suggesting ways to go, as long as you're not, you know, fully dependent on any of them being the actual path you take, but just as a way of exploring the possibility space around you at a given point in a creative process, I think clearly these are going to be an advance, you know,

a significant advance for creative people in that way. Steve, actually, one thought that I just had, is that in five, ten years, when we do this again and talk about the state of AI and GPT3 and Dolly, you're going to have an emergency in the middle of the podcast episode. You're going to have to go to lunch or take care of something. And I'm going to say, you know what, that's actually fine. We're going to stop recording. And then we're going to bring up Pod E, which is the podcast version of GPT3 or Wally. PodE will just finish the podcast for us. It'll be trained on my voice. It'll be trained on your voice. It'll be trained in our style. It'll understand, oh, they're talking about the frontier of AI.

So I'll just finish that conversation for them. We'll fire it off. The producer, Devin, we'll put it online. It'll go up in the morning, and no one will even realize that you and I didn't even do the second half of the podcast. That was PodE.

So I want to actually, I want to broaden this Zoom lens one more time, because as amazing as GBT3 and Dolly are, the rest of the AI frontier, I think, is utterly fascinating. And there's a couple important pieces of it that I'd love to

get your mind on. When you think about the fact that these technologies are tools for recombining existing information to produce creative outcomes, you think, well, where else can you use that? You can use it in writing, you can use it in art. What about in science? There is an antibiotic drug called Hallison that was discovered because MIT researchers used a machine learning algorithm to compare a bunch of different chemical combinations that might be likely to be a super antibiotic drug that could be resistant to E. coli. And the AI discovered the combination

that was, or several combinations that were most likely to serve that purpose. They discovered the superior drug through this AI search function. Another really amazing implication is AlphaFold. This is the project at Alphabet, the company that owns Google, which announced just over a year ago that they had developed a program that can predict the structure of proteins.

They trained it on all the protein structures that existed and kind of just like a really smart, you know, Gmail function that predicts night at the end of, it was great to see you last. It can do that, but for protein shapes. And by understanding the particular folding of proteins, we can predict exactly how the proteins work and we can design drugs to manipulate them. I mean, this is just extraordinary stuff at the medical frontier. Are there other examples that are most exciting to you about, about, you know, where we are in, in, in, in, in, in, in A.I. right now. Well, I think those are, those are great examples. And, and in a way, you know, they're, they're running in parallel to what we just discussed, which is that the AI is extremely good at generating new possibilities, which we then have to do some of the, you know, the legwork of figuring out what the actual, you know, useful possibilities are in that mix. So I wrote about this a little bit in Extra Life, which is the last book I did and the PBS series that we did. We talked about both those projects.

And if you look back in the history of kind of drug design, there are kind of like two phases to it. The first phase, which went on forever, which is serendipity. You know, you just, you stumble, you leave a petri dish out on your, you know, you're like, oh, wow, there's a mold growing in this petri dish. I wonder if that could be used. Oh, look, I invented penicillin, right? Like, that's, that was the technique forever. And then about 40 or 50 years ago, we developed this technique of rational drug design where it's like, oh, look, we understand a little bit more about kind of the molecular

chemistry of these things. And so maybe we could start thinking about kind of designing drugs from kind of first principles of how these molecules interact. And that, you know, like we got the AIDS cocktail out of that. And that was a big breakthrough. But what AI proposed. is this new approach, which is we scan through the immense possibility space of all the different compounds that could potentially be created that might have some, let's say, anti-bacterial

properties. And based on what we know about the ones that work, we take that, you know, the a million, you know, put the 10 million possible configurations and winnow them down to 30. And right now, the machines aren't good enough at then doing, a kind of a simulated drug trial. So we have to then take those compounds and test them in animals and testament humans and, you know, refine them from there. And so there's still that, you know, years-long process of figuring out if it really works that we have to do by hand. But the discovery process, which could have, you know, in some cases taken 10 years, could now be compressed down to, you know, 10 hours. And that's a huge advance.

The question is, you know, do we eventually get to the point where, for instance, you think about the, you know, the fundamental bottleneck, as you've talked about, we've talked about with the COVID vaccines was not drug design. Like, as everybody knows now, I think, you know, they design the Moderna, you know, spike protein approach in an hour. Like, they knew basically what they were building in an hour in, you know, in January of 2020. The problem was, or February of 2020, the problem was that it took time to. to one, scale it up, manufacturer, but particularly to run the three-phase trials, right? That was just something that had a time sync into it that was inflexible. Theoretically, five years from now, ten years from now, that we would build complex enough models of the human immune system and the human respiratory system that we would say,

okay, you know, here's a virtual version of this vaccine, apply it to these virtual humans, and let's run, you know, a million simulations of all the potential interactions that this drug or this vaccine could have. And we'll at least be able to, you know, maybe, you know, accelerate the process by you probably are still going to have some humans in some kind of phase trial, but you might be able to take it from six months to a month, which would save millions of lives with something like a pandemic, like the COVID pandemic. And so that's on the horizon as well.

I want to pause here and give listeners a sense of how all these ideas connect together. We've talked about Gmail autocomplete. We've talked about GPT3. We've talked about Dolly. Now we're talking about other technologies that could potentially predict new superior vaccines and drugs. There's a way in which all of these technologies are doing the same fundamental thing. They're recombining information to produce predictions. That's how you get typing.

It was great to see you last. Gmail Auto Complete. Night. That's how you get GPT3, predicting language, Dali, anticipating images, even these kind of drug technologies, predicting superior drug interactions with an immune system. What I want to do here is sort of take the dark turn.

I think we've talked about a lot of positive aspects of these technologies, but there's a lot of negative aspects as well. In fact, for all of these technologies, you can imagine them being used for ill. You know, GPT3 can be used for plagiarism. Dolly can be used for deep fakes. Drug technologies, like I described, that create antibiotics like Hallison,

could also make drugs or viruses that kill us. Talk to me a little bit about where your head is in terms of what the most negative implications are of AI that you're paying attention to. Well, I think those are major ones. I think the other one that obviously comes up lot is what it means for the future of work, right? You know, what, what, you know, I mean, I, one of the experiments I did with GPT3 was, I

said generate a lease for an apartment that is being rented for 12 months and, you know, cost a thousand dollars a month. And there's a sub, you know, I just listed the facts and it just generated this like, what seemed to my non-legal, legally trained brain to be perfectly legitimate lease with all the facts in the right place and, you know, and it did it in 30 seconds. So anybody, in fact, another thing that GPT3 does is it can code. So you can say, write me a JavaScript program that does this and it will generate the code. And they've kind of built out that side of it in the whole suite of other tools.

So just with the large language models, anybody who deals with very structured language-based information like a lease or a licensing agreement or code, there is a argument to be made that in 10 years, a machine will be able to do it faster with the same level of accuracy. So what happens to, you know, do those people just do their jobs more efficiently, but keep their jobs, maybe? Or do people lose their jobs because the machines can do it faster?

So that's the other, you know, question here. And so it, it, to me, I mean, the good news is, I think we are having a fairly robust conversation now while these technologies are still in the early stages. I mean, no one is really using a large language model in a commercial way other than in auto-complete. So this is not a technology that is out there in the consumer world. And yet already, you know, my piece is just one of many that have talked about these ethical issues and raised a bunch of these, you know, concerns about where this stuff is headed. That was not the case, you know, when Six Degrees and Friendster and MySpace were starting up the social media revolution, people were not talking about the potential downside of like an advertising-based model and how it might affect our polarization. And all those things were not being discussed.

You said so much there that I think is so interesting. Let me break my response down into two pieces. First piece is you reminded me of one of my favorite ideas in AI, which is called Moravex paradox. The idea that the hard problems turn out to be surprisingly easy, but the so-called easy problems should not to be surprisingly hard. So, for example, designing an AI that's a chess master sounds really hard. Turns out it's one of the first things we did in the 1990s and 2000s. On the other hand, like, you know, teaching an AI to behave like a child, walk down the stairs,

declutter a table, we still basically have no idea how to do that. The hard problems are easy. The easy problems are hard. The second piece is we, I want to lay out a, a. menu of ways to think about the negative implications of AI. And I have three categories here. Work, values, and profit. So first, work. You just mentioned this. It's going to be a little bit weird, I think, and surprising that a lot of these high status jobs, maybe lawyers, radiologists, jobs that require a certain kind of knowledge, recombination, scanning of language, quickly reading

images, these might be jobs that AI is surprisingly good at, faster than that. then it is good at, say, flipping burgers. Number two is values. We already touched on this, but I am just, I'm still so fascinated by the sort of AI culture war that's coming when we debate the values that we're encoding into our technology. And it's actually related to number three, which is profit.

Who's designing these technologies? It's Google. It's Microsoft. It's Facebook. It's open AI. They're going to make hundreds of billions, if not trillions of dollars with this technology. And that profit is going to be an amazing.

immensely important part of our conversation. So there's your menu. Work, values, profit. Where do you want to start there? Yeah, I mean, I think we cover it a little bit of the values question. It's incredibly thorny. And there's much more to be said about that. But on the profit side, the one thing I would add to that, that actually I didn't get a chance to bring up in the piece, but I think is really important, is when you think about, let's say Open AI or Google or whoever it is, ends up creating a large language model that truly becomes a major breakout hit. Maybe it kind of becomes the new replacement of a search engine. They solve the kind of reliability problem.

Let's just say that the true believers are right, and that's in the future. And that's an enormous, that's a trillion dollar innovation, right? That would create incredible value for whoever owned it. Now, what actually is creating the intelligence in that model? Now, on the one hand, it's the supercomputer cluster that is doing all those calculations, and it's the design of the neural nets and all that kind of stuff. But all that stuff is useless without that corpus of the World Wide Web and Wikipedia being fed into it. And so the intelligence is a mix of things created by the company that made the model and the internet, the collective intelligence, the internet that all of us made. And I think we, on some fundamental level, like, we have an ownership stake in that.

And it's, it is not correct. And this is an argument that has already been made a little bit about social media, right? It's a Geron-Leneer argument that, like, we're creating all this value by doing all this content for Facebook for free. We should participate in in some way. But I think it's even more true with this kind of artificial intelligence where you're really training it on everything that the world has put online. And these companies are coming in. sweeping all that up and saying, well, great, we can build incredible products on top of that. Like it's this natural resource that's just sitting in the ground somewhere.

But in fact, it was something that was created by human beings all around the world. And so I think that's one place where the profit side of it, we need to be way more imaginative in thinking about who created that value if, in fact, it does become as valuable as I suspect it will be. That's really interesting. And of course, in order to bring about a world, in which we have ownership over these technologies, due to the argument that we, by our contributions to the Internet, help to produce them, that will require new laws.

It will require new regulations, right? Microsoft is not going to come out with a trillion-dollar technology and say, you know what, we had to think, and this really belongs to all of you, so it's free. Like, that's not going to happen. The very last point, this is a point that Eric Schmidt made in a recent book that he co-wrote with Henry Kissinger and someone else called Age of AI.

It's about culture, the culture of a world in which AI is a more permanent fixture of our experience of reality. So he made this point recently where he said, what is it like to be a human in a world run by AI? Well, for one proxy of that, maybe look at Facebook. Like, Facebook has employed very intelligent machine learning engineers to develop an AI on that news feed

that maximizes engagement. And when you go on to Facebook, you are to a certain extent seeing what AI wants you to see. You could say the same thing for Google, maybe, ironically, there's search results are powered by AI too, but let's stick with Facebook for now.

Facebook recognized that outrage creates more engagement. High arousal, negative emotions create more engagement on social media. So there's more outrage on social media. There's more outrage on the news feed. Now, that's an incredible social experiment, right? That's a snapshot of a world where information that we see is partly selected by an AI trained on the knowledge of human engagement.

So if we have personal robots that are doing this for us, that are not just providing GPT3-style creativity inspirations, but also filtering the world for us. And they say, you know what? You know what Derek really likes? He really likes being outraged. He really responds very predictably and emotionally to outrage. I'm just going to feed him a lot of stories that make him super outraged. Well, then, you know, the AI world in which I am navigating is just swimming with stuff

that makes me feel sick to my stomach. To what extent do you think that is a reasonable fear that AI will sort of detoxify the informational landscape? I like the idea that in the future you might have a personal. robot, Derek, that would be like, did you see the thing Ben Shapiro said today? That it's just like sitting there instead of like doing your dishes, it's just designed to get you as aroused up as possible. Yeah, I mean, this is, these are epic questions. And to me, this is why, you know, to me it comes down to something we alluded to before, which is that we need to do a lot more thinking about the kinds of institutions.

that we have that develop and explore the possibilities of new technologies. And the model we have right now is that new technologies are large, with an academic world, somewhere in the mix here, they are largely kind of proposed and created and disseminated by venture-backed startups, you know, heavily concentrated in a very small part of the United States, a couple of spots in the United States and a couple of spots overseas by a very small number of people. And we have, you know, people in Washington and people in Brussels and a few other places who then, you know, years after these technologies have been brought out into the world, try to play catch up and say, hey, wait a second, hold on, that was maybe a bad move you made

with your advertising model. Could we rein that back a little bit? Or your privacy, you know, policies are really bad. We're going to ask you to change that. And that's not, I'm not saying this is somebody who's like, against regulation. I think we desperately need regulation. in this way, but the existing model doesn't kind of work fast enough or isn't there at the origin of these technologies. So we don't have a, we don't have a mechanism at the beginning of the process when we start to think about how these things could be used, where a broader polity is involved in that decision-making process. Regulations are always just playing catch-up with something that's already been unleashed in the world. And so to me, I think, you know, there is a

kind of a governance question here. And Open AI, the organization, behind GPT3 is the most interesting experiment in this and that they were founded specifically as a nonprofit. Their fiduciary responsibility in their charter is to the benefit of all mankind and not to their shareholders first and foremost, which is different from any other major, you know, big tech company. There are some other equivalents to this. They were smaller companies. They did introduce a for-profit arm because they just couldn't get enough money to fund what they were trying to do. And so they created this slightly complicated structure, which some people are, you know, suspicious of for perhaps for a good reason.

But it's still the decision of what to release, what safety guards to put around it, what values to use in training the model. Right now, that is all being kind of adjudicated by the shareholders and the board of an organization in San Francisco, California. Yeah. And this is why I actually want to close on OpenAI, because if you're You're right, and I think you are right, that governance and regulation will be too slow with AI as it is too slow with everything else.

We are unusually sensitive to the priorities, the organization, and the values of the companies that are making these products. We need them to be good. We need to hope that they are good. I am not here to argue that open AI is good or bad, but I am very very. very interested in the structure of this company. And I know that you're very interested in the structure of this company. You went out to San Francisco.

You visited them. You talked to them for this article. What should we know about this Willy Wonka factory of AI? What's important for the average person to know about them? Yeah. I mean, I think they began, on the one hand, you can look at it and you can say the people involved. You know, actually one of the founders was Elon Musk, who seemed. to be in the news for some reason.

Although he subsequently left OpenAI, we believe, because of conflicts with Tesla, which is also obviously in the AI space. Right. So Elon Musk was a part of the original team that came up with the idea for OpenAI, but as Tesla's AI department became more and more sophisticated, he realized he couldn't be simultaneously running competitors. So he's no longer embedded with the AI team, but he is one of its founders. But the rest of the core group, those there are people like Sam Altman.

Former head of Wycombinator. And Reid Hoffman was one of the people involved. This guy Greg Brockman, who was at Stripe. I mentioned this all just because these are significant figures in the Silicon Valley world. And so I think for some people, there's a very reasonable assumption to say, like, this is just another big tech. you know, if your default setting is that Silicon Valley-based big tech is the new big tobacco and it's just, you know, just a, you know, a seriously negative force in the world, then there's a reason to be concerned about OpenAI because it is the same people,

even though they've devised this different kind of structure. It's not a traditional venture-back startup or a public company. In my interactions with them, I feel like, you know, my personal assessment is that they are earnest in what they're trying to do. I think they put a lot of resources behind safety and trying to think through these problems. They've restricted the use of the software precisely because they're trying to slow down the adoption of these tools. It's the opposite of Move Fast and Break Things, the famous Zuckerberg mantra at Facebook. So on that side of it, I come away from my interactions with them. with what they're doing. I do not think while they have paid kind of lip service to, while they have

paid lip service to the idea of building tools that truly benefit all of humanity, they haven't built anything yet in place that would allow all of humanity to actually help make these decisions. And that is the hard part of the puzzle. If it ultimately comes down to all these decisions about how to release these tools and put them out in the world, if it's just, 20 people in San Francisco making those choices, then ultimately, I don't think they can live up to that, you know, kind of audacious goal. So. And we're talking specifically about the products that are coming out of Open AI, we talked about GBT3 and Dolly. These are free technologies that other companies can build businesses around. Is that right? That it falls to other companies to essentially choose to commercialize the technologies that open. AI is building?

Or does it work some other way? They're releasing them as an API, which means that you can build software on top of them. Dolly is a very closed API. Like you can't, you know, it's very hard to get access to it because it's brand new. Dolly 2 is. GPD3 is more open. Basically, to the extent that it's a business, the plan right now is that they're charging people for access to the API.

So if you want to build a business, you're going to have to pay the gas fee. this is an old Kevin Kelly idea, actually, about like AI being, you know, something that's just metered. You know, it's just like you get a certain flow of AI from, from, so you plug in the wall, you get some intelligence out of it instead of getting electricity. Every year we get our government ration of AI. Yeah. So they would, you know, to the extent that they now have investors through this kind of new arm that they created where they did raise some money, they would try to make some kind of profit to return that investment. But that investment caps out at 100x,

so that the investors cannot make more than 100 times their money on their investment, which seems like a lot. But early investors in Google or Facebook made 10,000 X, their initial investment. So by Silicon Valley numbers, that actually is a meaningful ceiling. So, yeah, and basically it's charging for companies or other organizations to build on top of the infrastructure of GPD3 and Dolly. That's the model. The very last question for you is what is it that you can tell us about what OpenAI is

cooking up that we could potentially see and be astonished by in six months one year? What do you think is going to be the next great reveal? Well, let me answer that in a slightly different way, which is that all the other companies are working on large language models as well. like Google and Facebook meta, presumably Apple. And OpenAI has made GPT3 way more public than the other ones. And so it's going to be very interesting to see what the next one that actually gets into kind of end users' hands in some form, what it's going to be like. But just the other just two weeks ago, right when Dolly 2 came out, Google released a paper.

that was using their new large language model, which confusingly is called Palm. We were talking about another approach called Palms. And in that paper, and people should look this up because it's amazing, they gave the Palm, large language models, a series of jokes. There were original jokes that had been written, you know, exclusively for this task.

They did not appear in the Internet anywhere else. And they asked the model to explain the jokes. and it's pretty astonishing. If you're trying to wrestle with a question of like, you know, are these machines synthesizing and summarizing existing information or are these machines in a kind of mindless way in a non-conscious way actually understanding the information? And to me, if you read the joke descriptions, and again, these have not been replicated. This is an internal paper that was released by Google.

We need outside researchers to replicate these results. but the explanations of the jokes are very sophisticated, and it's hard to read them and not think that there is some kind of emergent understanding that is happening there, understanding in quotes again, because it's not obviously a conscious system. And that, you know, that made me think, I really would like them to make that language model public because I would like to spend another three months messing around with what it's capable of.

So there's going to be, if, If you do anything with words and language, which is pretty much all of us, the next four or five years are going to be really interesting. You know, that makes me think. I know that there are critics of GPT3 that say that it can't think. All it's doing is recombining knowledge and spitting out predictions. But I do wonder, like, what if that's what thinking is? What if we're in the process of building a machine mind, learning that basically all we do when we're thinking

is recombining and predicting? One of my favorite economics papers is actually this paper called Recombinant Growth by the Harvard economist, Martin Weitzman. And I just pulled it up as you were talking. I'm quoting from the paper right now, quote, new knowledge depends on new recombinations of old knowledge. This paper's main theme is that the ultimate limits to knowledge

lie not so much in our ability to generate new ideas as in our ability to process an abundance of potentially new ideas into useful form. End quote. Here you have an economist saying that maybe the history of human innovation is just putting together existing ideas to create new ideas. What if that's all it is? What if a combination of preexisting knowledge multiplied by a prediction machine is all we've got, but we're just really, really good at it. And we have millions of years of a head start on these machines because we've been evolving for millions of years, but we're doing Darwin on, you know, warp speed in OpenAI and Google and Microsoft and Facebook, and we're just learning, wait, that's it. Recombinations, prediction. What do you think? You know, I think the thing

that people get hung up on here is if you think of thinking, if your definition of thinking is the internal experience of thinking, what happens in, you know, kind of the sentience, the experience of like bouncing around ideas in your head, then it's probably true to say that these machines are not thinking. There probably is no, almost certainly there is no internal experience of what it is to be like GPT3. But if you think of thinking as the process of manipulating concepts and coming up with new configurations of concepts that then can make more accurate predictions about what will happen next in the world, if that is thinking and intelligence in terms of the output, like what actually do you get out of the process, then I think

there is a path to that kind of intelligence in the near term future. Whether we would get to a point where you know, you would actually go to the AI as a kind of Oracle and say, how should we organize society or what should the tax rate be next year? I think that that's pretty fanciful. But in terms of, you know, going to an AI and saying, hey, I've got, you know, a set of ideas, give me some new combinations of them and maybe make some new associations for me with these ideas that you can make because your memory of the world and of everything that's ever been written is slightly more robust than mine is, that is a real kind of augmenting. intelligence that I think is a meaningful step forward.

And we shouldn't underestimate the power of that advance, even if it's not that kind of all-knowing, omniscient AI from science fiction. It's thrilling, scary, fascinating stuff. Stephen Johnson, thank you so much. Yeah, it's my pleasure. Thank you very much for listening. Plain to English is produced by Devin Manzi.

If you have a comment, a concern, a question, an idea for a future show, please email us at plain English at Spotify.com. That's plain, no space, English at Spotify.com.