How I Built This with Guy Raz - ICYMI... HIBT Lab! OpenAI: Sam Altman

Episode Date: February 16, 2023

Artificial Intelligence was once the realm of science fiction. But over the last several years, advances in machine learning and deep neural networks have moved us closer to a reality where computers can learn and solve problems independently, the way a human does. From art and music to medicine and politics, the potential applications of AI are nearly endless, and the technology just keeps getting better.This week on How I Built This Lab, Guy talks with one of the leaders in the field of AI development, Sam Altman. Sam talks about his journey from Stanford dropout and teenage entrepreneur to president of the legendary startup incubator Y Combinator and co-founder of the nonprofit OpenAI. Plus, Sam shares his hopes and fears for the future of AI and how his company is working to ensure it ultimately benefits all of humanity. See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

This podcast is brought to you by Squarespace. I talk to entrepreneurs all the time who are looking for a way to upgrade their digital footprint. Well, whether you're just starting out or you're scaling your business, Squarespace is the easiest way to build a great website that stands out. It's an all-in-one website platform that gives you everything you need to claim your domain, showcase your products, and get paid. Anyone can use Squarespace's cutting-edge design tools to build an online presence that truly reflects what makes your business special.

There are templates, intuitive drag-and-drop editing, and even an AI-enhanced website builder. Then, Squarespace's built-in analytics tools help you make smarter business decisions. Review website traffic, learn where to focus engagement, and track revenue all in one place. Looking to grow your business, Squarespace even offers fast, easy business financing

through Squarespace capital. Go to Squarespace.com slash built for a free trial. And when you're ready to launch, use offer code built to save 10% off your first purchase of a website or domain. Loans issued by Celtic Bank and serviced by Stripe, all loans subject to credit approval. This show is in partnership with Airbnb. This past summer, I took my family to Vienna, and it was incredible. We spent our days wandering the old streets, stopping for coffee and pastries, visiting museums, and just soaking up the history of one of the most beautiful cities in the world. And one of the things that made the trip so special was the home we booked on Airbnb.

It had tall windows, beautiful old details, and plenty of space for all of us. And being in that home on Airbnb, right in the middle of Vienna, walking distance from so much of the city, made it feel less like a visit and more like we were actually living there. Plus, taking a trip is the perfect time to host your space on Airbnb. your place with all of its personal touches and its amazing location could make someone else's vacation even better. Your home might be worth more than you think. Find out how much at Airbnb.ca.ca. Hey, so you've probably been hearing a lot about this artificial intelligence tool called chat gpt. It's basically a really powerful chatbot that can do things like

write essays, summarize articles, conduct research. It can even generate computer code. And while chat GBT is still relatively new and still has a lot of room for improvement, it has sparked a lot of conversation about what this kind of AI technology could mean for the future of work and education. Some universities and schools have even banned or limited its use, including New York City public schools, the largest school district in the country. Anyway, back in the fall of 2022, I spoke with Sam Aldman, the co-founder of Open AI. It's the company behind chat GPT and other AI tools like the image generator Dali. We talked about the enormous potential of these tools to transform society,

but also the company's approach to preventing some of the more sinister applications of this technology. With AI back in the news, we thought we'd bring you this conversation once again. Stick around. I think you're going to want to hear what Sam has to say about the development and future of artificial intelligence. All right, here's our interview. Hello and welcome to How I Built This Lab. I'm Guy Raz. So artificial intelligence is getting really good, really fast. For many years, most of us assumed AI was going to replace jobs in manufacturing in the service sector.

But what's actually happening is that machine learning devices are becoming much more adept at generating things we often associate with human creativity. You might have seen that piece of digital art that recently wanted to be able to be able to be able to be able to be able to. a competition in Colorado. Artwork completely drawn by a piece of software. AI software is getting really good at composing original music, even writing newspaper articles. And it means that in the not too distant future, machines might be able to diagnose diseases

or answer any questions we have or maybe even solve some of the world's biggest problems. One of the pioneers in AI technology is a nonprofit company called OpenAI. It was co-founded by Sam Altman in 2015. Sam is the former president of the legendary tech incubator Y Combinator, the accelerator program that helped launch Airbnb, Stripe, Coinbase, Storedash, and so many others. Today, Sam's on a mission to make artificial intelligence work in a way that has the maximum benefit for the maximum number of people.

OpenAI has made headlines with some pretty impressive AI tools, including GPT3, It's a language generator that's written op-eds and scientific research papers. And also, Dolly, which is an algorithm that can generate incredibly realistic images from any written description. But of course, there are lots of open questions about what happens if we can no longer control how this technology is used, which we will hear about. Sam Altman grew up in the St. Louis area, obsessed with computers from a pretty early age. I was a nerdy kid, that's true, but also like, that just the power of this idea that we were all going to be connected to each other, which I would say was evident by, you know, maybe like when I was like 10. Yeah. That, like, I was just like, this is like more important and more exciting and more fun than everything else going on.

And this is just, this is what I want to do. From what I gather, like a big kind of moment for you was getting that your first Mac 2. Yeah, Mac LC2. That was a really big moment. And how did you use that computer? Did you play games on it? I played games. I used AOL, which was sort of what I thought was the internet at the time.

And then a little bit after that figured out there was an actual internet out there, which was even cooler. I read like most of the encyclopedia on a CD-ROM. Yeah. I learned a program. What were you programming? In like basic or in? In basic. Yeah.

And what were you making? You know, like very little stuff, like print out all the prime numbers. like make a little interactive text game, stuff like that. But it didn't matter. It was just like the thrill of getting to make anything. Yeah. The excitement of like watching somebody else use something you made.

Yeah. It was just unbelievably exciting. Actually still is, but especially as a little kid. And so you described yourself as a computer nerd in high school, but was it clear in your mind that this was what you were going to do? Or were your parents like, hey, you know, you're really good at math and science. Maybe you should go to med school or something like that. It was totally clear in my mind.

My parents are great. They never, they never, like, pushed me to do something I didn't want to do. There was, you know, gentle suggestions of science or whatever else. But, like, it was very clear to me from a very young age, and I'm thankful for this. Because I know it doesn't happen to everybody that, like, computers were my thing. Yeah. Or technology, more broadly, was my thing. What was it about computers as a teenager that you remember feeling so passionate about?

I mean, was it a community of other people that you connected with, locally or even beyond where you lived in St. Louis? That was certainly part of it. I also was into like ham radio as a kid. And one thing that I noticed across both of them was the fact that you could like, A, there was interesting engineering and science problems and technology to figure out. There's like the fun of using a computer at all, especially back then when like they didn't work that often.

You had to like upgrade parts of them yourselves. That's just like a cool puzzle and technology is interesting. But definitely one thing that I felt across both of them is this ability to connect with anyone around the world instantaneously, which was incredibly appealing to me. The ability to sort of like learn anything, talk to anyone was just like, and still it's like pretty magical. I think we've we've all just now accepted that as standard. But if you rewind the clock like 30 years, things were very different. Yeah. I mean, I think the computer revolution is as big of a deal as the other.

great technological revolutions of the past, as big as the agricultural industrial revolution. But we've kind of just lived through it so recently that we don't talk about it as historic as it's going to turn out to be. Yeah, for sure. So when you set out to, I know you studied at Stanford and you studied computer science, so it was clear. You were very focused on entering this field. And presumably, you know, when you get to college, your thought is eventually I go worked for one of these big companies. Yeah. I picked Stanford because it was the best computer science program. I mean, it turned out to be like a wonderful place in a lot of other ways, but that was what I wanted to study. And my dream job after school was to work at Google. That didn't happen.

That didn't happen. While you were, while you were Stanford, you basically had a business idea and you dropped out. Yeah. Presumably with the intention of coming back. Absolutely with the intention of going back. One thing that's very cool about Stanford is they let you stop out and return sort of no questions asked very easily. Yeah. So you drop out to pursue this idea. It's your business called Looped. This is like 2005. So this is pre-Iphone, pre-apps. The idea was, hey, on your cell phone, because people had cell phones at that point, you could see where your friends were all at the same time on a map? Yeah. And also, you know, meet new people near you, find like interesting places to go or events to go to. And and so the idea was that like,

because I'm trying to go back to 2005 and I know I had, I definitely had a Nokia at that time. A Palm Trio was like the height of coolness in 2005. The Palm Trio. Yeah, right, right, right. But most, you know, the great, great majority of the world had like a Nokia something. But there were all of these like little mobile app stores. There's something called J2ME at the time where you could like make a little mobile app on a on a flip phone or whatever. It wasn't a very compelling experience. But it was sort of the idea was like as people lived more mobile lives, knowing everything in the context of the location would be super powerful. We turned out to be mostly wrong about that, by the way. I think it didn't really work for a bunch of reasons.

But one of them that is still somewhat surprising to me is even as the world has gotten very mobile, most people are sort of still at home or work all the time and there's just like far less spontaneity than it felt like to like a, you know, 20-year-old or whatever I was at the time. Yeah, I mean, it's interesting because on the one hand, I think most people wouldn't want to do that with other people. You can do that on an iPhone. But what they do do is they post on social media all the time. So we do know where people are all the time. Exactly.

So it's a different version of that. I'm not giving away the plot here by saying that loop ultimately did not sort of become what you envisioned. But it did bring you into this world that would become a really big part of your life. Y Combinator. You were in the sort of inaugural class of Y Combinator. Paul Graham, famously one of the main co-founders, I guess you met him and he liked the idea. But what's the story? Yeah. So I was like, you know, interested in this idea playing around with it. A guy that lived in my freshman dorm that named Blake Ross hosted something, I believe on Facebook, but memory could be wrong there now, that there was this new thing, and it was called the summer founders program at the time, and, you know,

people should check it out. The deadline was like the next day or something. I had followed Paul Graham online, was like a big fanboy. So I immediately applied with my co-founders, and they invited us out to interview. That was the first, like meeting the four YC founders, Paul, Jessica, Trevor, and Robert, that was like the first time I felt like, all right, I have like finally found the kind of people I want to be around. I didn't really know about the. I didn't really know about startup thing, you know, you, I heard about it as like a kid during the dot-com boom, but I wasn't really paying attention. I had met a few VCs on campus, but I could like tell I didn't like them. The world felt impenetrable. But I met the, I was like, wow,

these are my people and everything they say resonates. And even the other founders I met their interviewing. That was this immediate click for me of like, this is awesome. This is what I want to do. These are the people I want to be around. This is so cool. What were you doing at Luke? Like, were you actually writing, the software? I wrote the first version of it. Then I got pretty busy with other things. I think a lot of people assume they're going to keep coding, but hiring, managing the team,

my job ended up being like a huge amount of business development. I spent a lot of the time on airplanes flying around to the mobile carriers that were partners. But I was not able to code for very long. Yeah. So you're basically, this as a 19-year-old kid, right? It's a crash course. At one point, you raised money at a valuation of $175 million. You raised almost $40 million.

First of all, how did you learn how to become that person you had to become, to be in charge of a company with $40 million in capital that was entrusted to you? I had really great advisors. I think this is something that is still underappreciated about Y Combinator. People are capable of learning at a very fast rate. And one of the things that PG says that has always really resonated with me is you can get tough really fast. And that's sort of like one of the skills that is important for young founders. But if you jump into the deep end and you are like surrounded by supportive, thoughtful advisors, it was like interesting to me at the time how quickly I could learn.

And one of the most fun things about running YC later was helping other people learn very quickly. And just to clarify, when you say PG, you mean, of course, Paul Graham, one of the founders of Y Combinator. Yes. And I think this is one of the things that YC still just like people on the outside don't see it. There's this assumption that like, oh, if you get into YC, you're gold-plated, you have this like, you know, stamp of approval. That's why the companies do better. Like all the VCs are into it. But YC does so much that no other program as far as I know has figured out about how to get people up the learning curve really quickly. that it's a very big deal and it was super important to me personally and I think to many of YC's biggest successes it's been super important and the difference that you can make there

is huge. Yeah. So this was your life, looped was your life with support from Y Combinator for several years but ultimately it just didn't it didn't take off. It was acquired for I think for about $43 million. Yeah, it was acquired for like about what we raised So the later investors didn't make much money.

You earned a little bit, actually more than a little bit, you know, I think about $5 million from that sale. I earned some, but, you know, like this is another thing that is just sort of, I try not to think about. I made like orders of magnitude more money from angel investments that I made that I spent no time on the thing I poured my life into. Right, right. But after that happened, you were, in 2014, you were named president of Wycombinator. already by that point, why Combinator had kind of become this sort of legendary incubator. Yeah.

Tell me a little bit about sort of taking on that role. Did it feel scary, you know, to be, because again, you were still pretty young, leading this thing that had gained this kind of massive reputation. You know, it, this may be some sort of, like, weird flaw of mind. It didn't feel scary at all. It felt a little sad. Like, it took PG a while to convince me,

because I really wanted to do another startup. I wanted to prove myself. I was 28 or something at the time. I didn't feel ready to go retire and do a career adventure, which is really how I thought of it. I didn't have a ton of respect for the career path. And so there was like some sense in which it was an admission of defeat about, you know, I just can't run a company,

so I'm going to go do the easy job or I'm going to go do the retirement job or something. and it took me a long time to get comfortable with the idea of doing it. But I can't overstate how much at the time and how much now still I love YC, feel incredible gratitude towards it personally, also think it's just this incredibly good force in the world. And as I thought about the things that I wanted to do, I had this list from when I was like a college student of what I wanted to work on. It was like a long list. I realized I couldn't start all of those companies.

I also really was interested in pushing more investment in hard tech, deep tech, whatever you want to call it, which I didn't think there was much of happening at the time. Right, because of the time, WISI was focused mainly on like software, right? And you wanted, you wanted them to focus more on things like nuclear energy, self-driving cars, like actual physical things that you use. Not necessarily physical things, but I would just say like hard technology. Like there's a thing about Silicon Valley, which is it's, it has. not been that well geared towards the time and capital intensive science projects. So I wanted to do that, and I kind of was just like thinking about, you know, I had like basically taken a year off. It had been really fun. I had like learned a lot about everything. I had met a ton

of smart people. I had read like many, many dozens of textbooks. And there was like all of this stuff that I wanted to do. And I was like, you know what? In some sense, yes, being a VC is like an admission of failure. But in some other sense, what an incredible platform. I could like do whatever I wanted here. And I love YC so much. All right. So you became president of Y Combinator. And lots of YC companies have been on how I built this in the past, as I think you know, including Stripe and Airbnb and Reddit and Coinbase and lots of others. And I think maybe like a year in to that job, you co-founded this. nonprofit called OpenAI, which is mainly what you're focused on now. Tell me what was the

thinking behind it. What was the idea behind Open AI? So I studied AI as an undergrad, which was like dark times. Nothing was working. And then in 2012, deep neural network started to work. And not only did they start to work, it appeared that the more compute you throughout them, the better they got. Are you talking about like Watson and these these kinds of neural networks? The one that really kicked it off, I think, for the field, certainly for me personally, was the ImageNet result. This is the identifying images based on descriptions. Yes. So you would type in the computer horse eating an apple or whatever, and it could find them very quickly.

It was more like you'd show it an image and it would correctly say, this is a horse eating an apple. That's right. Or at the time, it would just be more like, this is a horse. And that was 2012. Yeah, it came out in my kind of like year off when I was just learning about things. and I looked at it with like great interest, you know, but kind of like it was one of many things that I spent time on that year. But I kept watching it and kept paying attention. And certainly, like of all of the hard tech dreams, AGI is the top of the list.

And just to clarify, Sam, when you talk about AGI, artificial general intelligence, you're distinguishing that from very narrow specific types of artificial intelligence? Correct. So, you know, we already have made good progress. with narrow AI that can categorize images or one task like that. But for me, AGI is a system that can learn and self-improve and create and reason about new information like a human does, this amazing general ability. And I think if we're able to accomplish that, it will be perhaps the most important invention in human history.

It will be the culmination of this phenomenal amount of effort all the way down the stack. and like the collective knowledge, accomplishment, moral progress, creativity of humanity into this one system that I think will be an even bigger revolution than the computer revolution we talked about earlier. We're going to take a quick break, but we'll be back with more from Sam Altman, co-founder and CEO of OpenAI. Stay with us. You're listening to How I Built This Lab. Welcome back to How I Built This Lab. I'm Guy Raz. And my guest is Sam Altman, founder of the startup loop.

former president of Y Combinator and now co-founder and CEO of the artificial intelligence company OpenA.I. All right. So Sam, in 2015, you and some other pretty well-known people, including Elon Musk and Reed Hoffman and others, form this thing called OpenAI, which is a nonprofit designed to do research around artificial intelligence. But beyond that, what was the purpose of OpenAI? What problem was it designed? to address. It was to figure out how to, and still is, to figure out how to build, deploy, and share the benefits of safe, general artificial intelligence. And it didn't seem like there was enough effort being put into this, given the stakes.

And so the purpose was looking at this thing saying, this is as important as anything we can imagine. And the upside is sort of unimaginable. the downside is catastrophic. And we look around the world and we don't see people taking this seriously. We don't see people caring. We don't see really any good safety research.

And so the thought was really like, can we help nudge us in a good direction? Now, AGI or at least powerful AI is like the hottest thing in the world. Everyone's interested in it. But what I can't overstate enough is at the time, we really got started in January of 2016. So it's like, you know, six and a half years ago.

At the time, people thought we were crazy. Like, good researchers say you've totally discredited yourself by talking about AGI. Most other people were like, this is, you know, 100 years away or more. There's no idea about how to do this. And we just kind of like got cracking on it and had a couple of huge discoveries, some good engineering. And now here we are. Why did you found it as a nonprofit? I mean, initially, I know that you have a for-profit subsidiary now, but why did you originally make a

nonprofit. I think this technology really does deserve to belong to the world as a whole. It's going to have such a profound impact on all of us that I think we deserve, like we globally, all the people, all of humanity, deserve a say over how it's used, what happens, what the rules are. We deserve to share in the benefits. We deserve to all have access to it. And that was why. I'm like very pro-capitalism, obviously. I think capitalism is great. But I think AGI is sort of an exception to that. In a well-run society, I think even if we rewound the U.S., you know, 50 plus years, this would absolutely happen by the government. Right, because DARPA would have done it, as they did with the Internet and so much of the other technology we use every day.

Yeah. I think most people, most of us, have different views of what artificial intelligence means, right? And I guess ultimately it's whatever our imaginations can conjure up. But for some people, it could be a version of like, you know, humanoid robots from, you know, Westworld or machines that can detect disease in humans based on, you know, hundreds of thousands of data points. It can mean a machine that can beat the best gamers or chess masters. I mean, all of these things. But, but of course, that's how we see it now. When you talk about artificial general intelligence, what do you imagine in your mind, what do you think it actually could do? The first thing that I think it'll be and where people will use something like, wow, that is real intelligence of some sort. What I imagine in my head is a gigantic skyscraper full of computers and we all text it like we would text a coworker or a friend or whatever.

And increasingly powerfully for whatever you want, it can answer it. So you can say like, I need life advice or I need like great medical advice. and it'll help you with that. Or if you're just trying to like search for something to find information, it can help you with that. It can look at your email and say, you know, here are all the important emails that came in overnight. I've drafted a response to all of them. If these look good, hit send and you'll be happy with all of them. You'll want to hit send.

Already today, you can say, I'd like this art, please. Can you generate an image that looks like this for me? Can you write this code for me? Eventually, now this will require more compute, but if we, society allocate to compute to that, you can say, like, could you go off and find a cure for cancer? And it can spend a lot of its compute cycles doing amazing scientific progress like that. But I imagine that we get to a world where each of us has sort of like a, what feels like an AGI companion that we are talking to all day, that is helping us be the best version of

ourselves, learn, be efficient, be happier, and we'll all experience that way. And then globally, we will say, okay, this is like very expensive, but we're going to dedicate it to a few huge projects. You know, maybe we'll have it help us figure out how to identify like better diplomacy between countries or the cancer example I mentioned or figure out how to do carbon capture. But when you have this like very superhuman intelligence, I think the harder challenge is not talking about limits of what it can do because that's sort of like beyond our imagination, but how to conceptualize it. And how to like think about what it'll look like at the beginning. I know that you went out originally and attracted a lot of sort of donors, right, a lot of capital to begin research. And one of the first really public things you essentially worked on was to train an AI to play the game, the video game, Dota. And it's a really complex game, right?

And not to bury the lead here, but your bot won. I mean, beat the reigning world championship team. Yeah. Why was that, why is it sort of the ideal place to start in AI development? You need a really good environment. At the time, we kind of had a lot of people on the founding team and early employees that were good at reinforcement learning. And reinforcement learning, we should explain. This is when you feed a lot of information to the AI.

The AI is basically taking random actions. Yeah. And then you're giving it a signal for, okay. that worked or that didn't. Yeah. But the thing we learned with Dota is you can do it over a very long time horizon. So there's like, you know, hundreds of action steps a minute. There's thousands of things you could choose for each action step.

And games can last, you know, the better part of an hour. No one thought that was within the realm of RL. That's just sort of too long to go without a reward signal. And we did experiment with some intermediate ones. But what it turned out is by playing gigantic amounts of time, hundreds of thousands of hours of the game per day, and having the agent sort of play against itself. So it has to get better and better to keep winning. Just by trying random actions and reinforcing the ones that have a positive impact, the AI system can learn to beat the best humans at this game, which was not obvious to us, quite remarkable given the size of the action space. How is it different than like deep blue beating Gary Casper?

of in chess? There's 16 pieces in chess. The search space is just not nearly as wide at any given mood, even if you do have to think pretty deeply about each one. And so chess, people were able to beat even without neural networks. That just turned out to not be as hard as we thought. But again, this is like the complexity and the action space of this game is just beyond what I myself. did not think we were likely to succeed at the project all the way when we started.

Sam, given that you guys are taking a big swing here, right? You are looking to become the preeminent, groundbreaking AI research organization in the world. I'm curious how you go about the sort of the series of steps you take in deciding what next to focus on. I know there are multiple things you're focusing on, but I'm talking about the things that we hear about. Yeah. So, for example, that, you know, Dactyl was something. that came out next. And this was basically, and people listening can see it. It's a video of it. It's a robotic hand or arm. And it solves a Rubik's cube with one hand. I mean, it's cool. It's very cool. But I'm wondering why that was the next big challenge you wanted to solve. How does that move the needle just slightly forward on AI technology? Well, it turned out to be a mistake.

I mentioned that Dota was a great environment. It turns out that robotics are a very hard environment. Robotics is hard, and it's not hard because the machine learning is hard. It's hard because the simulator is bad, the robot breaks. You can't run it that fast like you don't have like hundreds of thousands of robots, like you have hundreds of thousands of CPUs to run a game on. And so we thought it would be interesting because robotic control seemed very interesting to us the time and still does. But what happened is it was too hard of an environment and we ended

it up pausing our robotics work. Someday we'll come back to it when the robots and the simulators get better. It's clearly, you know, it would be great to have, but it turned out to be hard in the wrong kind of ways. The way we pick projects to work on is like not as exciting as everybody would hope. You know, I think there's this belief that you have a bunch of people like sitting in a room picking this like brilliant secret strategy. We just sort of like run our own reinforcement learning algorithm. We do more of what works. We follow the technology. We realize that it's like very hard to sit in a vacuum and predict where everything is going to go. Most people who have tried that have been catastrophically wrong and it's very smart people trying it. The technology

just kind of like surprises all of us. And so we pay close attention to what's working. We put more effort into what's working. We then put effort into figure out how we can deploy it safely and beneficially. And then we go do the next thing on top of that. But it's very much like driving on a country road at night with headlights on. You can always like see before you have to make a turn, but you'd have no idea what the next few turns look like. Right. And that's a question, right?

Because if you are building a product, right? If you're building an autonomous vehicle, you know what you're you're building towards. But you're really not building towards anything in particular. You're building towards something that we humans can't even imagine, right? And that's the thing. Like, it's, I guess it's a matter of just figuring out how to develop technology that can think like a human beyond how a human thinks can process information in highly complex ways. Yeah. You bring up a great point. The other company I'm involved with is this nuclear fusion company called Helion. And there are a lot of similarities between the two companies in terms of like a very hard

scientific and engineering problem. But the biggest difference is like what to do, what success looks like, what's going to happen on the other side of this, what problem to go solve next. You can look at it backwards or forwards. It's very clear. It's all hard, but it's very clear. There's no uncertainty about what to go after, what's going to happen, what the company needs to do. And another nice thing is like it's basically all upside. There's none of the

harder issues to contend with. At Open AI, we just don't have any of that certainty. And one of the things that has been like a little bit surprising to me is how difficult it is to get advice about how to run a project or a company in the face of such uncertainty. It hasn't happened a lot of times. And I am like wistfully envious of the helium world because it's, you know, it's still super important and super great. But it's so much clearer. And we just have to like turn over one card as we go. Yeah. Sam, one of the one of the things that you did in 2019 controversially was to form a for-profit wing of Open AI. And Originally, it was kind of this heralded nonprofit.

You know, hey, we're going to make this technology available to everybody. We're not going to compete. We're going to collaborate. All of our research will be available. We want to be a responsible, you know, steward of this technology and to encourage responsible use. 2019, you form a for-profit version. Plenty of people, including some people on your, you know, involved on your own board who were like, wait a minute. What?

You know, why? 100%. totally get why people don't like that kind of change. What happened is really quite simple, which is when we started OpenAI, we did not expect massive scale to be as important as it was turned out to be. And by 2019, we realized that and that the amount of money we were going to need to succeed at the mission was beyond what we could raise as a nonprofit. Yeah.

You know, way more than 90% of that for compute, but also buying expensive data sets, compensating people in competition with Google who can, you know, pay very huge salaries. Yeah. And we realized that like either we had some way to like dip into the power of capitalism and the ability to get the resources we needed or we were just going to be irrelevant. We did, you mentioned the government earlier. I had forgotten about this until right now. We did also just see if the government wanted to fund us.

They definitely did not. Yeah. And there was no other source. of capital that we could figure out. And so we wanted to preserve as much as we could of the specialness of the nonprofit approach, the benefit sharing, the governance, what I consider maybe to be most important of all, which is the safety features and incentives. So like, for example, we have this one thing called a profit cap where our employees and investors can only make a certain fixed amount of money with their equity. And then beyond that, all other

profit is distributed as fairly as possible with the world. And I think, you know, at the time, that was like another thing that seemed crazy to people. But I think that's going to be really important because I think, and you already see some people thinking this way, it's going to look at some point like, wow, you can generate close to infinite wealth with AGI. And people who have equity, even if they start out not wanting that, are like, hmm, you know, maybe that sounds better than I originally thought. We also have something in our documents which says if we need to for safety, the board can just

totally wipe out everyone's equity value to nothing. And we have something called the merge and assist clause, which says similar to that, if another effort's ahead and we want to avoid a race condition, we can just shut down and merge with some other effort. These are things the normal company wouldn't be able to agree to, while still being able to access the capital we needed, which it was clear by that point would just be far, far greater than we ever thought. Yeah. You did bring in a lot of money from Microsoft. And as a result, they get an exclusive license to use some of your technology in their software, also criticized by some. But, you know, I mean, again, you got to bring in the money. So was that, in your view, a compromise

you had to make? And if it was, how does it impact the openness of open AI? I mean, if Microsoft gets exclusive use of some of the technology you're developing, does that in some ways undermine the original vision? You will see us open more technology. over time, we just like to be pretty sure on the safety front. First of all, the most powerful model in the world, as far as we know, of an available language model, is still the one we created two and a half years ago that is available in our API. And that just tell me what that is again? Oh, it's called GPT3.

It's like a powerful language model that people develop on top of. Right. And that's available. That's open source. Anybody can use that. It's not open source, but anyone can use it. I got you. Okay.

We could open source. it's someday in the future. It's certainly something we'll consider. But there is a thing about the API access, which we like, which is when we know less than we'd like, when we're in the fog of war, if we just publish the weights of a model on the internet and then we realize like, there's actually a safety issue here, we can't take that back. It's done. It's like a one-way door. We also cannot put any usage restrictions on it after we open source it. When we deploy it via the API, we can take it back if we find that there's a real safety issue. But more importantly than that, we can enforce usage restrictions. We can make changes if unsafe things are happening.

So let's let's kind of break down GPT3 for people who don't know this, because obviously, as you know, how I built this is a generalist show. We talk about chocolate chip cookies and artificial intelligence on the same show. And from what I understand, GPT3 can do a lot of things, but there's a famous example from 2020 in the Guardian newspaper. They published an op-ed. They essentially asked GPT3 to write an op-ed called Why We Shouldn't Be Afraid of AI and published it. And it's pretty interesting. I wasn't convinced by the argument the AI made, but that was pretty cool. So this is one, because when you talk about safety, right, you're talking about people misusing this technology to, for example, make, you know, have AI create fake information or

manipulative content? You can imagine a lot of things, but the disinformation manipulative content, I think, is high on the list. GPT3 at this point, we're more comfortable is relatively safe, although I think there are risks related to misinformation in some cases. But we'd rather be too conservative than not. Like at some point there will come a time, one, two, three orders magnitude. Who knows what, how much more powerful than GPT3, where I think a language model can really have a huge disinformation effect on the world.

And again, kind of the traditional stance of the tech industry is just like, yeah, you know, launch the product and think about it later and deal with the problem later. But like, we don't want to push a button through a one-way door. Or at least if we do that, we want to be incredibly confident and careful about it. Yeah, we're talking about five, ten years away, I think, more or less. Is that fair to say? Yeah, I would say, you know, sometimes these things take longer than you think. But let's say like in ten years, I think there will be powerful digital intelligence in the world. We're already seeing this with music and art.

I think a piece of art just won an award and it was very controversial. But we're already seeing relatively high quality content produced. Yeah. I should say we did put in, we asked GPT3 to write us an intro, an overview of Sam Altman's early life for how I built this. How was it? It gave us two versions. The first one says, Sam Altman is the president of Ycombinator and co-chairman of OpenAI. He was born in 1985 in Cambridge, Massachusetts.

He attended high school at the Phillips Academy Andover, where he was a national merit scholar. The second one said, Allman was born in 1985 in Berkeley, California, raised in the town of Lodon. Salos, attended Los Altos High School, where he was president of the Science Olympian team. Both days are wrong. He grew up in St. Louis and did not go to either of his high schools. But it does have other information in there about Open AI. So, I mean, you know, these are coherent sentences. Like, you can, you can imagine that a human could have written them just just got the facts wrong. Yeah, that's a big problem with the current version of language models is we haven't trained them yet to try to, like, be helpful and to verify that what they're saying is accurate. They're just. just try and sort of sound like coherent text. We have a lot of ideas about how to fix that, and my guess is in the next couple of years will make significant progress there. But I think that is the current biggest single problem with language models is, as someone

famously said in there, they're very convincing bullshitters. And it makes it very hard to use for a case like this. Now, as you mentioned, an easier problem that has worked very well, we have a system called Dolly that will take text and generate images. Really cool. You could put in like a giraffe dancing ballet on top of a hypersonic jet, and it will generate that image. Yeah. And people love it. Like people who never cared about AI before that I went to high school with or whatever

are like now totally obsessed. And, you know, I talked about it and it's very fun. But it's like it, we talked about this idea of the technology gradient and kind of doing what works. Like we still have more work to do on language models and that's hard. but in a world where we don't need perfect accuracy and we want creativity, man, the image generation thing is so cool. Yeah.

We're going to take another quick break, but when we come back, more from Sam Altman, co-founder and CEO of OpenAI. You're listening to How I Built This Lab. Welcome back to How I Built This Lab. I'm Guy Raz, and I'm talking with Sam Altman, former president of Y Combinator and co-founder of Open AI. Sam, I wonder, I mean, we're talking about all. all these cool things, right?

And we've talked about a lot of cool things. Now, here's the thing. I remember back in, must have been 2010. I interviewed Mark Zuckerberg at a public forum at the Computer History Museum. And the optimism about the future of social media was just so unbelievably clear in his mind, that it was only for the benefits of the world,

that it was going to be this amazing thing that will bring people together, We're all going to be part of this single global community. There's only going to be one version of who you are, the same person at home as you are online, and that it's going to open up all these channels of communication and cooperation. And of course, that vision was not realized. I know you're a really smart guy and you think about these things. And so you know that all of these wonderful things, you can answer questions for us.

It can solve cancer maybe. It could give us a diagnosis. It could identify the best doctor. that will happen and then there's going to be other things too. So let's talk about some of the other things and what keeps you up at night about those things. You know, I've had three recurring bad dreams over the course of my life. When I was in high school and college, I had this one about like missing a test. It's sort of evolved when I started my startup to like thinking in the dream that I was like still in school and I was missing stuff because I was running the startup.

after the startup, I had this recurring bad dream for a while just about failing, which was just incredibly shameful and very, very tough. You know, startups are great when they're working, but when you're failing, it's quite brutal. And the current one is about, like, what if we're wrong about all of this? I think there will be, like, good and bad with any new technology. And I can accept that as long as the good is order of magnitudes more than the bad. And that is truly what I believe will be the case. I think the world is going to be incredibly different, but there's no stopping that now or ever. And I think we will be able to responsibly deploy current technologies.

And I think one of the advantages of our nonprofit structure is we do a lot of things differently than a normal company would. But for me, the real question is can we solve the existential issues? And I am more optimistic about that than I've ever been before. I think a lot of the work we've done around alignment and safety is like promising. And we have like a plan that I believe in. But, you know, technology is impossible to predict. We can make high confidence guesses. But as you said, a lot of smart, well-intentioned people thought that social media was going to be like an unequivocal good for the world.

And it is good in some ways, but it's come with some real negatives. We try to learn every lesson we can about what's gone wrong there. And I am obsessed with this idea of incentives being a superpower. And if you can get the incentives right in an organization, then I think you avoid a lot of the sort of things that cause well-meaning people to do things that maybe later that wish they hadn't. Yeah. But my fear is that there's just an unknown unknown that we totally missed. and technology or social pressure or something else is going to push us in a different way. Again, I think we're getting a lot of things right now.

You know, we deploy our technology slowly and cautiously. Like we're willing to piss users off to go slowly. But as the systems get much more powerful, the challenges become more and more unprecedented. Yeah. There are a lot of intelligent people on planet Earth. And some of those intelligent people are bad actors. I mean, you look at how bad actors have deployed social media or digital technology to steal money from people's bank accounts or crypto accounts or, you know, just these are not dummies. These are intelligent criminal networks, right, that have figured out how to use a variety of tools that are now available to them to do real harm. And that's going to happen with this technology, right? There will be really intelligent people who figure out how to get an AI to, create a new virus and deploy it very effectively. Like that could happen or how to, you know,

build your own very powerful explosive. Now, that being said, do we just accept that that is part and parcel of being human, that there are humans with bad intentions and there are humans with good intentions and that it doesn't matter what we do with technology because some people will use it for good and some people will use it for evil? And you just cannot prevent that. I mean, what you said is true. Some people will use technology for good and some people will use it for bad. But it doesn't mean we don't do anything. It means we work as hard as we can to have the technology be as useful for good as possible and as difficult to use for evil as possible. And it is in some sense, like somewhat one-dimensional, like more powerful tools, let bad actors do worse stuff. But I think, think it's like very defeatist to just say. And so there's nothing that can be done. Like our strategy, and it's as you touched on, been controversial, has been to deploy the technology in a way that maximizes the good and minimizes the bad. And, you know, there will be lots of technology

where it's turned out we're too conservative. And actually, we can just make it more open to the world. But I'm confident there will come a day where the world will say, like, it's good that opening eyes on the conservative side with new stuff. Most of us interact with AI already multiple times on a daily basis. Like when I type an email and Gmail, it makes suggestions, which are often very helpful. And so there is some of that already that we use, you know, Siri is a version of this. But let's say in five years from now, based on what you are seeing and what you guys are working on, right, how do you think the average kind of person might be? might use the technology in five years from now?

I think it is just going to kind of be everywhere. You know, when the apps were on the iPhone launched, it would be a big deal for companies to talk about being mobile companies or their mobile strategy at that point. And at this point now, it's just like every company is mobile, and you wouldn't think about it. It would just be, like, ridiculous not to have a mobile app for most companies. But no one goes around talking about being a mobile company.

And I think the same thing will happen for AI. There will be a layer of intelligence provided by us and others that's just everywhere. And as the systems continually get smarter, everything you use will get smarter and smarter as well. So there will be like, you know, there will be the direct usage where we talked earlier about this idea that you'll chat back and forth with this like super intelligent assistant that will help you with everything. But also in everything else you use, you'll just have a very high expectation that it's really smart. So you mentioned that you kind of like the Google suggestions in your email. Those will look quite trivial, and you will expect every morning to wake up and for Google to have detailed, thoughtful replies that you yourself would be proud to come up with to every email in your inbox. Which, by the way, I'm thrilled for because if I never spent another minute in my inbox again, I'd be really delighted.

Yeah. You know, it's interesting because I think for a long time, there was just this feeling that AI and automation would replace like menial work or, you know, factory work, which is still can. I mean, you know, you could make the argument that there's no need for a barista because a machine could do everything that somebody making a cup of coffee, not as beautifully. Doesn't say hi to you, doesn't smile to you, doesn't give you that minute of human connection. Or is it nasty to you either, right? Or that. Depending on which coffee shop you go to. But now it seems like something different is is happening, which is AI looks like it could replace more and more creative types of jobs,

jobs that are quote-called white-collar jobs, maybe even before it replaces so-called blue-collar jobs. Absolutely. You know, the strong consensus five, ten years ago, was that first of all, the AI was going to come for the blue-collar jobs. Second of all would be the less sophisticated white-collar jobs. Third of all would be the very high cognitive load white-collar jobs like a computer programmer or a mathematician or whatever. And then last of all, and maybe never, because maybe this was like special and human-only, would be the creative jobs. And if we look at it now, it appears relatively clear that it's going to go in exactly the opposite direction. And there's a lot of things one can take away from that, but one of the most important ones

is predictions about technology roadmaps are hard, and a lot of very confident experts get them wrong. So my observation is just that, like, this was hard to predict. It seems fairly clear now. With the benefit of hindsight, it also seems fairly obvious now, but everybody was confident and wrong in the other direction.

Robotics are really hard, as we talked about earlier. Yeah. So, I mean, what does that mean in practical terms of? Right now we're talking here in 2022. There's a labor shortage in the United States, right? There's just that are not enough humans to fill all the jobs available. And there's a big crisis that many companies across the border, not just in the service sector, the blue collar jobs, but also in technology companies. So on the one hand, you can imagine, oh, this is great because, you know, maybe some of these jobs will be taken over by, you know, artificial intelligence. intelligence, you know, devices or software, whatever it might be. But this labor situation won't be around forever. There will be higher unemployment rates. And so what happens? I mean, do you, do you, I mean, if this technology is, is going to become so good, which it will be, what happens to the, to the humans? You know, there's a huge amount of pole clutching about what, what's going to happen when AI replaces all of the jobs and what we're going to do. And I have a few thoughts about that. One, I actually think we're in a humongous labor shortage, like much more than it seems. There are many people who love their jobs and derive great meaning from it, you know, but I think those people don't recognize it's a privilege that most people don't have.

And most people would love to work less or not or do something completely different. But we have sort of this like pay-to-play world right now where you have to pay to live. And I get why some people think that's important, sort of, in theory. but I personally really don't think that's the society we should strive for. I think like traditional work in the way we think of it should be optional and you should be able to do less of it. I really love work. It's my hobby. It's my passion. I think it's great. But I know a lot of people don't and a lot of people would choose to spend their time in a very different way if they didn't have to. I also think a lot of people who like work would clearly like to work less. We're seeing this in this post-pandemic world where a lot of people are like, oh, I'd go into the office two or three. days. I'd really rather only work two or three days, too. I have a lot of other things I like. No,

I really don't want to go back. And so I think the labor shortage is actually like much bigger than it seems. I also think like a lot of things that we do were understaffed even before the pandemic, but certainly like going through an airport now or going to get medical care now, you like really feel, you know, this is not what full employment would look like where I just like walk in, everything's ready to go all of the time. Or where like every student has like a one-to-one teacher to student. That could be amazing too. Who knows what. So huge labor shortage now also. We've seen this with every other technological revolution too. I don't know what the jobs of the future will look like, but I am confident that human creativity, desire for status, desire to

do new things and to like accomplish, that's not going to go anywhere. The economy and society will look super different. That's for sure. But we have always worried about this. We have always found something new to do. I do accept that this time it may be different, even though you're never supposed to say that. Maybe if we figure out intelligence, that's just very unlike anything that's happened before. But I still would never bet against our desire and ability to find very new things to be busy with. Sam, if that is the case, right, and we're kind of getting out of our lane here, which is fine, we should be. But our species, right, homo sapiens, we've been around for 330,000 years, more or less.

And really only, you know, kind of in the last 30,000 years have we started to widely spread around the world. And previous, our relatives, you know, the other hominin species lived a long time. Yeah. 1.2 million years, 1.3 million years. And then another species replaced it. We replaced the other ones to become the kind of the dominant homin species on the place. planet. We're talking about developing something that is more advanced than humans. Are we talking about, you know, a form of evolution? Are we talking about the next species that replaces homo sapiens?

I get the analogy, but I think it's basically incomparable. I hope that human homo sapiens are the end of the line for evolution in a survival of the fittest sense. And that from here on, even as things change quite rapidly, we are much more thoughtful and deliberative. And unlike the previous process, this time around, we can reflect, we can debate, we can build this system in a certain way. This is like very different than a random natural process. It's very important to me personally, and I assume to you and hope a lot of other, almost everyone too, that humanity is in charge of the future of humanity. And the systems that we build, how we choose to deploy them, what we all collectively want out of the future, that is a discussion that I think we need to start now and have a lot of societal coordination around. But I think the analogy, although it sounds very similar, is like different in almost all ways that matter.

this is not going to be just this like chaotic, stochastic process playing out, but we get to make all of the decisions about how this is going to work. While we're way out of our lane and talking about the very long-term future, I think if we end up in the us versus them framework, that's like some varying degree of bad. And the most positive long-term futures that I can imagine involve some degree of emerge. It doesn't have to be the crazy sci-fi, like, you know, plug something into our brains or upload ourselves or whatever. But in a minimum, I think it needs to involve something like, you know, this like AI alter ego that we're talking to all day that observes our whole life that understands us. Its goal is to like extend our capability and will.

I think those are the easiest worlds to see it being very positive. Yeah. Sam, you're not yet 40 years old. what is my guarantee here that you don't become like Robert Oppenheimer one day and say this thing that I was part of was terrible. I regret it. I have that book on my desk. I'd look at it every day. I mean, I can't give you a guarantee on that, right? But I can tell you I will work as hard as I possibly can and get the best people around us to make sure we don't end up there. This is going to change almost everything. This is going to just be a seismic shift. And the way we get to the best version of the world is to have as broad input on where we want to go, what we want to happen as possible.

And the way to get there is for a lot of people to take this seriously, understand it, direct some of their precious and limited attention towards it. And we really want to drive that conversation forward. That's Sam Altman, co-founder and CEO of OpenAI. Sam, thanks so much. Thank you. I enjoyed that. Hey, thanks so much for listening to How I Built This Lab. please do follow us on your podcast app so you always have the latest episode downloaded. If you want to follow us on Twitter, our account is at how I built this and mine is at guyraz,

and on Instagram I'm at guy.org. If you want to contact the team, our email address is hibt at ID.wondery.com. This episode was produced by Chris Messini with editing by John Isabella. Our audio engineer was Neil Rouch. Our music was composed by Rumpteen Arablewey. Our production team at How I Built This includes Alex Chung, Carla Estevez, Casey Herman, J.C. Howard, Liz Metzger, Sam Paulson, Carrie Thompson, Catherine Seifer, Josh Lash, and Elaine Coates. Neva Grant is our supervising editor. Beth Donovan is our executive producer. I'm Guy Raz, and you've been listening to How I Built This.