[0:00:02] Do you have a preference? There we go. It can only get better from that grand entrance. There we go. Well, we're so thrilled that each of you could join us today. And let me offer a warm UND welcome. I'm Andy Armacost. I serve as the president of the University of North Dakota. And today we have the special privilege of having Greg Brockman return to UND. Thank you.
[0:00:37] It's honestly so great to be back in North Dakota. Absolutely. And Greg is from Thompson, North Dakota. His folks are here in town right here, physicians. And your mom, of course, served here at UND and currently serves in a medical capacity here at UND. So there's a close connection between you and this campus.
[0:00:59] So Greg is the co-founder and president of OpenAI. And OpenAI, of course, has been at the leading front of the development of generative AI tools, and most notably ChatGPT and DALI and many others. And so it's just a thrill to have him here.
[0:01:16] One of the goals of OpenAI, of course, is to create a general purpose AI and really looking at the benefits to society about how we can use these tools to make life better for all of us, all of us human beings. And so we're so thrilled. ChatGPT, by the way, was the fastest growing app in the history of apps. And I think you had about, what was it, a million users in...
[0:01:38] Five days, right? So it was a meteoric launch. And not what we expected either. We called it a low-key research preview. We thought it was going to be like a couple hundred people, mostly our friends. And you got a million in five days. We got a million in five days. So we actually have these little stickers now that say low-key research preview and have a computer that's on fire and everything's going down. That's awesome.
[0:02:04] And so there's tremendous growth in the attention that it's garnered since its launch. And I, of course, think about a year prior to the launch of ChatGPT, November 22 of 2022.
[0:02:15] I think it was November 1st of 2021, there was a book that came out by Eric Schmidt, Henry Kissinger, and Huttenlacher. And they talked about the age of AI and what it means for us as humans and the challenges and the incredible opportunities that it presents. And so that was a prescient book. And then lo and behold, a year later was the launch of ChatGPT. So Greg has a long history of looking at not just AI, but the ethics and the philosophy behind AI and how to
[0:02:44] How do we make opportunities exist for society and for human beings? So we're thrilled to have you here, and we'll do a series of questions from me, but then we had also asked a number of students for their questions as well, and we'll go through those too. Sounds great. You ready? I am. Okay, first question. So tell us about your journey. I mentioned Thompson, North Dakota. Tell us about your journey from Thompson, North Dakota to where you are now. Yeah, so Thompson was a really
[0:03:13] Wonderful place to grow up. Just super safe. Like our doors didn't even really have locks. And I think that it's just very different in many other places in the country to kind of have the same kind of just like sort of idyllic lifestyle. So I actually was going to school in Grand Forks, which eight mile drive away. And so it was very rare that people would come over and
[0:03:41] Be able to hang out with people after school. So I had a lot of time on my hands. Time that I spent exploring my own interests. Got very interested in math competitions. Got very interested in doing research, some of which I did at University of North Dakota. And I also got very interested in playing computer games.
[0:04:01] Which maybe many people can relate. But I was very lucky that my parents actually were very supportive of these interests. And our dad, when I was in sixth grade, started teaching me algebra. And that was kind of the first step towards this sort of independent path that I ended up taking. And throughout middle school, I was doing independent study. And then high school, I was taking calculus in freshman year. And then I was out of...
[0:04:27] Out of classes at my high school. But fortunately, there's University of North Dakota right down the street. And so I took most of my classes at UND. I actually did three years worth of coursework here. And that was something that really helped me just delve deeper into what I was excited about. And I think that is maybe the biggest theme for my life and my career so far, has been to just kind of just try to take a path that's just different from other people. And it doesn't have to be the specific one that I did. In fact, it kind of shouldn't be, right? That if anyone is interested in
[0:04:55] Just finding what they're passionate about and just not following the beaten path. It should look unique to you. It should be something that is authentic to you. And so I graduated, and I went to Harvard initially. And I was there for a year and a half, transferred to MIT, dropped out of MIT to go do a startup. And the thing that really got me was when I started programming.
[0:05:18] With math that I've been really into, I was into chemistry, all these things. It's these very abstract thinking. There are these moments where it clicks. And when it clicks, then you kind of understand something and you solve this math problem or you come up with an answer to the chemistry problem. But with programming, you actually have a system. You actually have something that works. And anyone else can use it. They don't have to understand the details of what you did, just the intent, the vision that you had. And I was like, that's what I want to do. And so I actually spent all of my sort of formal university years
[0:05:46] Trying to do startups. And most of them are very bad. You've never heard of them, for good reason. But from each one, I learned another thing not to do. And then I met three people working on this payment startup in Palo Alto called Dev Payments at the time. Pretty terrible name. You've also never heard of it. But I was like, these are the kind of people that I just want to be around, I want to learn from. And so I dropped out to go just work with them.
[0:06:14] No matter where you're going, if you're dropping out of Harvard, your parents are going to be a little bit worried. So they were worried when I went to MIT. And then if you leave MIT, no matter what you're about to do, they're also going to be extremely worried. I actually sent one of our cousins to come check up on me to make sure that this company actually existed and that I was able to eat. I didn't actually have a bed at the time. I was sleeping on the floor doing the whole startup thing. But it was a really wonderful experience. I really learned how to build. And while I was at this company, which I renamed to Stripe,
[0:06:43] Which, you know, maybe people have heard of. I started to watch deep learning from the outside. You know, AI had always been the thing for me that maybe one day I'd be able to work on, but I wasn't sure when or how to contribute. And you could see, just like if you read sites like Hacker News, you could see something new in AI happening every day, deep learning every day. But I had no idea what deep learning was. And so I met someone working in the field, and he introduced me to more people working in the field. And I kept getting introduced to all my smartest friends from college.
[0:07:11] And that was a sign for me of, aha, I think that there's something real going on here. Maybe it's time to contribute. And so that was part of the impetus for me to leave Stripe and to start OpenAI. Yeah, and we were comparing notes backstage. And some noted professors of mine back at MIT were involved in traditional, called optimization techniques, which was my field.
[0:07:32] And they too, probably in the 2013, 2014, 2015 timeframe, had said the same thing. But there's something magical about deep learning and machine learning in general. And lo and behold, here we are today. And so let's go back to UND. And what are your memories of being here? And we have to give a huge shout out to Ryan Zier, Professor Zier. Where is he? Right over there. There he is. Who one day was sitting in his office and Greg comes walking by and says, hey, I think I'd like to study something here.
[0:08:03] Or that's my version of this story. Tell us about your version of this story and how it began. That's basically accurate. So again, I'd finished up the math at Red River High School in ninth grade. And that was actually a great experience. I was really into math competitions. I was just like, math, this is the thing. I want to study so much of it. And in middle school, I had to independent study classes because I didn't have a car. I couldn't go to the high school. But sophomore year of high school,
[0:08:31] Driver's license. So I was able to actually drive from Red River to UND.
[0:08:37] And I had a setup where I just took three classes at Red River and then whatever I wanted at University of North Dakota, which was wonderful. So the first class I took was set theory and logic with Dr. Zier. And I was definitely the smallest person in the class, but I had a great time and people were happy to work on problem sets with me and all that stuff. And I just felt very welcomed. And I took a bunch of other classes. I took organic chemistry with Dr. Smoliakova. I took philosophy.
[0:09:06] With Dr. Weinstein, I believe. And that was actually a really wonderful class because it was philosophy through film, Western philosophy through film. And so you actually watch a lot of movies and you have a good excuse for it. But it was just so interesting. There are all these little facts that you never think about. For example, one that really stuck with me is that water in Western symbology is chaos and rebirth. And so if you're watching a movie and it's raining,
[0:09:30] And the characters are not in the rain, that usually means chaos. And then when they actually go into the rain, that means rebirth and it's kind of the end. And if you watch movies with that lens, you're just like, huh, yeah, that's once again, it's like the rain, it's like everyone's upset at each other part, and then everyone goes out into the rain, and it's like, okay, this is now everyone reconciles. And it's just funny when you realize there's so many, you think there's so many degrees of freedom in our society that everything is made new, but actually everything's connected. Everything is just a slight deviation from everything that has come before it. And I think that realization that there's both
[0:09:59] Like, to create something new is both the synthesis of you have to understand the things that came before you and the context that you're in, but you also need that little spark of creativity to specialize it to your special case, to think of what is the thing that no one else has, what are the connections no one else has made. So that's something that really stuck with me, and it's a very, you know, it's not a technical thing, but I think it's actually really underlined a lot of what I've done. Nice. And in your short-lived stay in college, you had three degrees that are an interesting collection. Share those with us. Well, so I thought when I was here that...
[0:10:29] I was going to go and do a math, a chemistry, and philosophy triple major at Harvard or MIT. None of those ended up happening. I have zero degrees. Sadly, my wife is the educated one in the family, so we appreciate that.
[0:10:46] But just that mixture, the philosophy addition ties to what you talked about in Professor Weinstein's course and what you were exposed to there as well. So back to AI and the choice to kind of delve in. So you were at Stripe. You said, hey, something about deep learning is something you really want to pursue. Tell me more about that motivation about pursuing AI. Yeah, so for me, the beginning of the AI journey was actually after I graduated from high school. So I took a year off.
[0:11:12] Before jumping into these colleges for my short-lived stints. I didn't mean that as an insult, by the way. No, no, not at all, not at all. And I went abroad. And actually, in high school, I'd been taking Latin. By the way, anyone here take Latin? Very few. I recommend it. I think you learn, once again, you learn a lot of the underlying structure about why the world is the way that it is, at least our part in the Western Hemisphere. But I...
[0:11:40] Senior year, our mother said, you have to take a real language. So she was of the opinion of everyone who didn't raise their hand. And I was like, fine, I'm going to take Russian. Because it felt to me like in Western Europe, lots of people speak English. And Russia felt like just a place that I didn't know much about and I wanted to learn about. And so I actually took the year off after high school and went to Russia for four months. And there I, once again, had lots of time.
[0:12:07] It was very self-directed, and so I started to read about programming, because I'd never really done it. I was kind of curious about it. And I read the 1950 paper by Alan Turing, where he talks about the Turing test. By the way, how many people know what the Turing test is? Raise your hand. More hands than Take Latin. So it's the idea of how can you tell if a machine is intelligent? In 1950, Alan Turing, who's an amazingly brilliant computer scientist, proposed the idea that maybe what you could do is if you...
[0:12:37] If you talk through a chat system so you can't see who you're talking to, a human, and you talk to an AI and you can't tell which one's which, then you can say that that AI is intelligent, that it actually is smart. And he actually even laid out an idea for how to solve it. He said, look, you'll never program an answer. It's just too complicated. You can't write down all the rules. Instead, maybe you could build a child machine that can learn like a human child. You give it rewards and punishments when it does something good or bad. And maybe from that, you could actually solve this test.
[0:13:06] And when I saw that, I was like, this idea of you could actually build systems that can solve problems that you yourself do not know the solution to. You can't write down the rules, and yet you can build a machine that solves it. I was like, that is the coolest thing I've ever heard. Because there's so many challenges that we know we're not really going to be able to solve. We haven't made progress on them in thousands of years. Or new challenges. Think about climate change. How do you solve that? It's this complex coordination, technological, societal problem.
[0:13:34] Extremely hard, and so if we had better tools, if we had some way of building systems that could solve problems, help us solve problems that we could not on our own, like, that is the way. Now, the only problem is, nothing worked. When I showed up in Harvard, I actually went, did the same thing that I did with Dr. Zier. I went to a professor, I said, hey, could you, could I do research with you in natural language processing? And he showed me these, like, parsed trees, and I was like, okay, this is clearly not gonna scale. It's like, you're still writing down the rules. But 2015, that is when
[0:14:03] You started to see that, huh, there is a general purpose approach that is actually knocking down the walls between departments, whether it was image recognition where it started, machine translation, all these different fields. You have 40 years worth of research, people are throwing it out, replacing it with neural nets. And I was like, this is very interesting, and I just want to help support it. Nice, nice. Which ultimately, a couple years down the road, led to the founding of OpenAI and the co-founding. And did you know that the technology
[0:14:32] That you were building would have the impact that it has initially. And tell us about that vision that you had when you founded OpenAI. Well, I'd say yes, and even more. I think we are still very much in the early days. And that's a really sobering thought. That a lot of when we started OpenAI 2015, you could kind of, if you were in the weeds, if you were really sort of talking to the people that were doing it.
[0:14:58] You could see the shape of what was coming, right? You're like, okay, it's like these methods that are very general, but still wasn't necessarily something you would use yourself. It wasn't a consumer product. It was still kind of like a little bit, you know, there's always this S-curve from pure science to pure engineering, and we were starting to enter it. And, you know, we spent a lot of time, because a lot of the ethos of OpenAI is not just let's build the thing, but it's think about the impact, right? Like, I think this technology is so desirable, it will be built, with or without us.
[0:15:28] But what we can do, what we can all do, is help steer it, is help decide how we want it to integrate with society, how we want to work with it, how we want this to integrate with our lives. And so in 2015, we spent a lot of time really thinking about this. You take it really seriously. You sort of condition on your own success, and you think about what must you go through. And we really had this picture that, yeah, you know, that you're going to go through a lot of these, like, cool demonstration systems, and we did that. You know, 2018, we solved a competitive video game called Dota.
[0:15:56] You're going to go through a very commercial phase where this is like the most interesting new technology. Crypto was occupying that slot for a while, but hopefully we are delivering more lasting substance. But things are going to get even more serious, right? You think about how we organize the world and how...
[0:16:19] You think about impact on jobs and just like what it even means. Like if you move to a post-scarcity society, just how do you make sure that everyone has not just enough, but actually everyone has just like lots, right? Not just universal basic income, but universal high income for everyone. And especially in a world where you have just like a very different sort of relationship to where cognitive labor is done than we have today. And so there's a lot to think through. And ultimately, you think about the balance of power between different nations. But even beyond, if you build
[0:16:48] Incredibly smart systems. And we're not close yet, but we might be closer than people think. At some point, the technology itself is something you have to worry about.
[0:16:59] We've all seen the movies about the crazy AIs. And I think actually the scenarios that people think about are probably not the ones that will happen. But I think it is important to think about each of these risks of people sort of misusing the technology, of technology being integrated in society in a way that just somehow is just bad, and the technology itself becoming a risk. And so we need to kind of navigate through all of that. And I think that the most important thing is to capture the benefits along the way. And so, by the way, how many people here are chat GPT users? Have tried it at least once?
[0:17:28] All right, so decent show of hands there. More than Latin, yeah. Yeah, exactly. I'm glad we're more popular than Latin. That is good. But that is, I think, this system is the least good it will ever be. It's only going to get better from here, and I think that there's going to be step functions each year in terms of what we're able to do, not just in text, but across every single domain of AI. Well, so as we saw, many of the people in the room have used ChatGPT.
[0:17:55] For many, this is the first time they've knowingly had this type of interaction with an AI or a generative AI tool. And so when you look out over the next 10 years, I think you started answering this already, but the next big jumps, and I'm particularly interested in knowing, right now AI is really crafted around specific use cases. In other words, I want to look at something within medicine or cancer drugs, and there's an AI that's tuned for that. What about the steps towards...
[0:18:24] And artificial general intelligence, which is what many consider to be the kind of the end-all, I don't know if it's the end-all or not, but the ultimate goal of human-like intelligence with these machines. So the mission of OpenAI is to ensure that artificial general intelligence benefits all of humanity. That's what we're really here for. It's nice that we have a consumer product that has grown so fast, so many people use and love, but that's not our mission. It is a tool towards our mission.
[0:18:53] That we want to not just build the technology, but actually also integrate it with society. And that was something that was actually a very hard decision to make. That there were big debates about maybe what you should do is build this stuff in secret, and then one day you have something very powerful, you have time to make it safe and test it in the lab, and then you push go. And that was always really scary. The idea of it just doesn't seem legitimate to do such a powerful thing outside of society should choose where we go.
[0:19:20] And I think that the fact that everyone is being educated, everyone has their hands on this technology, is probably one of the most important steps towards our mission. And I think that at a technological level, the way to think about it is, and again, all this stuff is so hard to picture because AI, I think, is like this incredibly surprising field. If you look back at most of the thinkers about how will AI look
[0:19:43] No one really predicted it. No one predicted something like ChatGPT. People were thinking about, well, we'd write down all the rules, and we'd come up with some really smart algorithm. Instead, this picture of actually compute, the computational power seems incredibly fundamental here. And it's not all you need. Compute data algorithms are kind of the three pillars here. We're going to keep putting more compute in. We're going to keep coming up with better algorithms. We're going to keep cleaning the data better and kind of getting more representative data of the world.
[0:20:12] The way to think about it is that ChatGPT is this system that basically knows everything, but it still makes these very non-human-like mistakes. It makes these reasoning errors that you never would. And if it has two or three steps of reasoning, sometimes it can do it. But once you have 10 steps, it's going to fall off the rails and then say things that are not accurate. And I think, what does it mean if you can solve that problem, if you can actually have a machine that can have coherent thoughts and that you can just go off and think about a hard problem?
[0:20:41] And I think that that is where the promise really lies. That that's what we really want. We want to be able to present it with a hard material science problem. You want to be able to present it with even a hard coordination problem of just like, hey, we want to get everyone to the table to talk about this issue. Who should we call? What should you say to different parties? What does it seem like this person really wants? All of these deep things that feel very deeply human, but we're actually not that good at. That it's actually, you think about how much better
[0:21:09] We could be the better coordinator of a world. We could be if we had better machines to better tools to empathize with each other and really understand even just crossing language barriers and cultural barriers that are just hard to do unassisted. And I think that that's the place that we're going. And I think that it will help a lot with technology development. I think that today we're talking about thinking about how to use these tools in the classroom. I think it's actually a very important question too of how do we all
[0:21:37] Sort of become better humans and learn more and leverage these tools. But yeah, I think that the shape of the technology will become very different. And I think it will be very different at each step. So kind of expect each year there's going to be something a little bit surprising. And I think in 10 years.
[0:21:53] Like, that it's really hard to put a bound on where we might be. Like, I think there is a decent chance, you know, at this point, I think people are definitely saying, you know, within, it's within the Overton window now to say that within 10 years, AGI is, you know, it's not 100% guaranteed, but I think it's actually quite probable. And by the way, I think that one mistake we made in the early days was thinking of AGI as the end goal, right? That it's basically like everything is focused towards this one moment, and then either success or fail, and that's it. But instead, I think it's a continuum, and it feels way more realistic to me.
[0:22:22] Nice. And I know there are many predictions about when this happens. I think Ray Kurzweil predicted 2040, and there's been others. So it'll be interesting to watch these step developments that you describe as well. Yes. Actually, the Ray Kurzweil thing I think is so interesting, right? So I think Kurzweil actually is maybe the thinker who has been the most predictive of what was going to happen. So in the early 2000s, he was like compute.
[0:22:44] That's the thing. And everyone else was like, this makes no sense. So yeah, because he was drawing all these curves. He was like, OK, once you have enough compute to simulate a human brain, then you'll have AGI. And everyone's like, why? Who cares if you have a fast computer? You need to figure out how to actually program it. That's the hard part, right? And now, basically, people have said, OK, fine, Ray was right. We'll figure out the programming part. And he said 2029 for the Turing test, which when I think of that, that's kind of AGI level of a machine that really you talk to it, and it's fully human level.
[0:23:13] And it doesn't have this weird non-human mistakes. And 2045 for the full-on singularity, which is what you think of as what we call superintelligence. And that is going to be, I think, something that humanity should approach in a deeply coordinated fashion. We've been talking about this. Many world leaders, the need for an international organization like the International Atomic Energy Agency.
[0:23:35] And I think that if we can actually come together as humanity to build this kind of system, then I think we actually have a shot at something just like truly unimaginable. I was going to ask about balancing the risks versus the benefits to the world. And I think you just hit on it. So you do see a role for some kind of cross-governmental cooperation to make sure we keep things within bounds? Absolutely. The funny thing is that a year ago, and believe me, we've been having these kinds of conversations
[0:24:05] Since we started OpenAI, since really even before, the idea of international coordination would kind of laugh you out of a room. People were just like, everyone's in it for themselves. There's just no hope to actually coordinate. Government doesn't know what they're doing. All these things that are just standard tropes. And again, it never really resonated. It always felt like the world is not a stationary place. People update fast. And you can see since the release of ChatGPT, it used to really be hard for us.
[0:24:32] We spent a lot of time trying to outreach to various governments, and it was kind of hard to get them to pick up the phone. I mean, understandably so. They've got so many other issues they were dealing with. Think about the pandemic versus talking about some crazy AI thing. But it's been very different this year, that I think that every leader realizes that they need to get smart about this for their own country and for humanity, and that it's become actually a very...
[0:24:59] People are really latching on to this talking point of saying, hey, it's not just going to be every country for themselves in some sort of deadlocked race. I think that we need some sort of coordination. And of course, we shouldn't be naive. I think that everyone has some amount of thinking about their own citizens, and they should. But I think that there's also
[0:25:16] As you start to understand the power of this technology, the potential power, you realize that we do need to work together. We do need some sort of international guardrails and oversight. And you think about the kind of data centers. We're going to build basically these machines that are just these giant machines bigger than anyone's built before. And those are things that there should be some sort of international oversight for. And you actually can do it. It's not going to be that many organizations that will be able to build these things. And so I think that there is just this non-stationary equilibrium.
[0:25:44] And people kind of look at the way that things are, look at how everyone's thinking today. But what is kind of our responsibility, what we think about is, where will it be? Where is the ball going? And trying to build for that. And so we look a little bit crazy in the short term. We fortunately look a little bit less crazy now than we used to. We still look a little bit crazy. Don't get me wrong. But I think that time will tell.
[0:26:06] So let me pull back from the geopolitical context and to college campuses, right? So you have a room full of students and educators and what are you thinking about in terms of the impact of AI on how we educate future students? Yeah, I think that two things. First of all, one of the most, maybe the most important message that I want to get across today is one of empowerment, right? That we are not
[0:26:33] We're not deciding the future, right? We are building tools, but how do people use these tools? Where do we want them? Where do we not want them? That's something that we can't decide on our own. And I think that it's been really amazing hearing so many stories from different people about how they're using ChatGPT, right? And I think that the first reaction many people have when they see new technology is ban it. It's bad. Not unique to AI at all, but you look at New York City public schools, for example, and that was the first thing they did.
[0:27:02] And now it's unbanned. Now they're really leaning in. And I think that figuring out how to navigate that, that's something that I've actually been really, really encouraged to see. And so I think that it's a little bit like we're all going to be entering a world with AI as a tool. It's a little bit like computers. It's just like, who has not used a computer in the past day, including your cell phone, any other computer? If I see a hand, I'll be shocked. I think AI will be like that in a really deep way.
[0:27:30] And so, yeah, you need to be not just using it. You need to become an expert. I was actually talking to a friend who's a professor who's saying that they were trying to figure out at their school how to deal with students who might be using ChatGPT for admissions essays. And her kid suggested, actually, I think that you should admit the people who use ChatGPT the best.
[0:27:53] And it's just like this different perspective of not just thinking about, well, we used to work this way, and now we have this tool that kind of threatens it. And there's some of that. That's important. But I think that the real winners, and I think this actually can change even the balance of which institutions are the most thriving, like where do all the students want to go, is to think about what are the things that we couldn't do at all? What are the things that were totally impossible you couldn't even dream of, and now you can do with ChatGPT? And so I'm hopeful that...
[0:28:20] We are providing something that unlocks opportunities. And in previous interviews elsewhere, you've talked about AI as a tool that can essentially serve as an always-on teacher or tutor that can understand and perhaps even motivate a student. Well, if this happens, what's the role of the traditional educators? So, yeah, I think that it's a great question, right? And again, there's a question of some amount of...
[0:28:47] Direct replacement and some amount of new opportunity that was totally impossible. And I think about the fact that everyone has that one teacher that they remember who really inspired them, motivated them. We've talked about Dr. Zier. I think he definitely did that for me. But imagine if you could have that 24-7 for any subject you wanted and that you have this AI that you have a long-lived relationship with in the sense that it
[0:29:11] It understands you. It really has seen your learning progression. It kind of gets your style, that it kind of knows that like, hey, you really should be doing your homework now. Whatever the things are that you just simply could not find humans to do. And there's so many studies that show, there's like the Two Sigma study showing that personalized education actually is like basically the one thing that actually moves student outcomes. And it's like there's so much opportunity. And so I guess the question is, well, as a teacher with
[0:29:40] A lot of students, there's zero chance you're going to be able to be that teacher for all of them. And now if you have a bot that you're able to work with, too, and you're able to present it to the students and work with that, how do you bring that into your classroom in order to help you do things you could not do otherwise? And I think maybe this is an important thing. So we work with Khan Academy, for example.
[0:30:02] And that they took our same technology that powers ChatGPT and built a system for schools. And they're actually that you have the students and you have the teachers and the parents all as users of the system. And that's something that, again, we're not able to go and just build that ourselves. It requires a lot of expertise and a lot of thinking. And so I guess it's like you don't just want the raw technology. You want to build these layers on top that integrate it and give it to the right people with the right sort of oversight in order to make sure that you're getting the benefits. Nice, nice.
[0:30:28] We have a series of student questions as well that I'll go through. But first, my final question. Actually, I defer to ChatGPT on this one. So I'd like to say, if you don't like the question, tough. This is your doing. So if ChatGPT could experience emotions, what do you think it would feel when it successfully helps someone? And conversely, what would happen when it makes a mistake?
[0:30:56] Well, I certainly hope that it would feel great pleasure at doing a good job and maybe some anger or displeasure or frustration at not. And by the way, I think that the question that I've actually thought about a lot, first of all, I think we actually have no real answer for do these systems have internal experience? Do they really feel? They can say that they feel, but does that really mean that they do?
[0:31:18] And we haven't made much progress on that with humans. We kind of believe that other people can feel stuff because we can feel stuff, and we're like, you look pretty similar to me in terms of all your physical machinery, and so probably you have internal experience as well. But we don't have a mechanistic, oh yeah, here's the specific part of the brain. But if you just bear with me a little bit, one philosophical argument I think is interesting, or sort of mechanistic argument, is maybe what's going on is that
[0:31:45] Emotions, internal experience, these actual feelings are convenient computational shortcut. Like maybe it's, you know, we have these neural nets that are kind of finding the best way to solve a problem. And maybe the best way to solve a problem is to have an agent that when it does a good job actually feels good feelings, right? And then it's just like, you know, it doesn't have to like sort of write out the like, oh, I should probably feel really good here. Just like, it just feels it, right? Just like a kind of a shortcut. And so maybe actually you could get some emergent.
[0:32:11] Intelligence, where the pleasure is hooked up towards the positive and that the displeasure is hooked up towards the negative. The thing that would be really scary is, well, what if they're just reversed? Why should they be aligned in the right way? Oops, I put a negative sign in that equation. Exactly, right? Yeah, yeah, yeah. Definitely, you don't want that bug. Yes. So I think that this, by the way, will be the most fundamental and interesting and important question in the long run, because what you really want is you want to know. I mean, first of all, you really want to know if every time you train your neural net,
[0:32:39] Are you actually causing a pain? Is it a good thing? Thinking about model welfare is a thing that may become important. And then we as a society have a choice. Maybe we don't want to build systems like that, or maybe we really do. Because I think that being able to expand to the cosmos with beings that really feel and think and have a great time in that cosmos, that seems like a pretty good long-term future. Nice. Well, let me turn now to the students' question. Our students always ask the best questions. So here we go. This one comes from...
[0:33:08] Dev, who's a computer science graduate student from Grand Forks and a 2014 graduate of Red River High School. And Dev writes, what are the important lessons you learned in your journey to get where you are today, coming from a small town in the Midwest, and what unique perspectives have you gathered along your experience in AI? Well, first of all, great question. I'd say, first of all, my life algorithm, in some ways, professionally, is to...
[0:33:33] I think one thing I do somewhat uniquely is I get really excited about the upside of things. For AI, I got so excited. I was like, wow, if this thing could work, it would be so great. And then I actually go explore it a little bit. And so, for example, I went to that professor at Harvard and said, hey, could I do some research? And then I also make sure there's a very fast out if it turns out that the excitement is not quite warranted. And there, it was just like, nope, never mind. Not going to do that. And no harm, no foul is all good.
[0:33:59] But I did the same thing with many startups, right? That all these startups I did throughout college, you know, I kind of worked them for a little bit and kind of realized like, oh, this isn't quite the right way to do it. The first one that I did was actually these two seniors that had this idea that was so secret that they would refuse to tell anyone and they begrudgingly told me and then, you know, they built a whole team of like 10 people and I was the only programmer. And I kind of realized like, these people don't know what they're doing. I was like, I don't know what I'm doing. And I was like, huh, this is probably not going to succeed. But again, it was an experience where you kind of realize by going through it.
[0:34:29] OK, you actually really need, unless you yourself are ready to be the visionary, you need someone who knows what they're doing. And so that was kind of the next thing that I moved to. And so I think just that algorithm of trying to get excited about what could be and then making sure that you're not irrevocably in it until you're sure, that's something that I think was very important. And I think that I brought that to AI. I think that we're very much in the, this one I feel like,
[0:34:57] We're definitely doing it. I'm excited about the upside. I also don't think there's much of a way out at this point. I think that I'm very, very much committed to seeing this through. Nice. And when you think about promising applications of AI in the near future, we'd love to know more about this. This question comes from Lexi, who's a finance and information systems major from Las Vegas. Yeah.
[0:35:18] It's very easy to come up with promising applications because basically everything is on the table now. It's like such an exciting time. I think that for me, one of the things I'm most excited about is actually applications within medicine, which again, naively, if you rewind a couple of years, I would have said those would be the last applications because medicine is such a, you make one mistake and it could be someone's life.
[0:35:40] But actually, it's turned out to be one of the easiest applications. And not that you just say, oh, replace a doctor, and now you put in AI, you're all good. No, you don't do that. It's the opposite, right? That you think about how much time physicians spend doing things that are not doing the hard part, that are not trying to figure out what disease you have and what medicine you should take and things like that. It's like all this paperwork and transcription and all these things.
[0:36:02] And AI is helping with all of that, that it's integrated into Epic, which is the biggest medical record system. It's integrated into a bunch of different places in the doctor's office. And so I think that being able to remove the busy work from the most important professionals and to let them do more of what they do well, that's something that is already happening. And the thing that I'm most excited about is eventually being able to do new technology development.
[0:36:29] To be able to come up with answers to problems we don't know the answer to. And the next question was actually about potential of AI in medical practice. This is from Odell, who is a nursing major from India. And I think you hit on it, but now maybe we expand that discussion to talk a little bit about the use of AI in development of therapies and new drugs and so forth. I know there's been some remarkable progress there as well. Yeah, I think this part's really interesting. Because the thing about the ChatGPT angle is it's very good at text data.
[0:36:57] It's like a bunch of language and to be able to parse through things. And actually, there's a bunch of medical applications you wouldn't think of, like for example, taking a bunch of medical records and trying to parse out whatever medical billing code it's supposed to be, things like that. Which again, why are doctors spending their time on that? But I think that then moving to things like new drug discovery, that there are systems like AlphaFold from DeepMind that we're able to map.
[0:37:21] The folding of every single known protein that I know someone who's thinking about how to take DNA sequences and be able to apply this technology to it. And so I think that just like you can think of the language of biology as something that's a total mystery, right? Like we really don't speak it very well. It's just like there's so much like going on there. There's so many connections that it's kind of hard to keep track of in this thing. And so we're starting, it's like the bar in some ways is low.
[0:37:47] In order to make progress and be able to solve problems that you couldn't before. So it's still early days. I don't think we're quite at the point yet where drugs have been discovered that you absolutely would not have been able to do without either some sort of machine learning techniques or any deep learning techniques. I think five years from now, it'll probably be the opposite. Nice. And in medicine, you recognized, wow, this involves humans and life. Let me stay on the human side as well. And this question comes from Madeline, who's a commercial aviation major from Sequim, Washington.
[0:38:16] And the question is, do you think that artificial intelligence will degrade our value on man-made art such as poetry or painting? No, I don't think so. I think that things will change. I think there's no question, right? And I think that one thing we've already seen is that more people are artists now. More people can come up with a poem. And I've seen this with so many users of DALI, you know, our system that turns text into an image.
[0:38:44] That there are people who had some picture of something that they wanted. And sometimes it's as simple as, you know, someone's thinking about decorating their home and that they want to be able to visualize different options for it. I think that there's a lot of these kinds of applications that you wouldn't think of that someone, there used to be this barrier that the tools were so hard to use that you'd have to become an expert before you could use them. But at the same time, I actually think that the value of, you know, now I can make all these beautiful pictures, but someone else can too. And so,
[0:39:15] The value of those maybe is like less special if they're just kind of like, oh, I just kind of like didn't even write a prompt and the AI just kind of came up with a thing and here's the thing now. No one wants to look at that, right? And it's actually very interesting that we've seen this lots of times where people...
[0:39:28] Tweet out some picture that they made. Tweet out some experience they had with ChatGPT describing it. And no one wants to read it. It's just not interesting. What people want is they want that human connection. That if you're going to stare at some art, and sometimes you do want to just stare at beautiful art because it's beautiful or because it was made by an AI. But sometimes you really want to know that the person who's singing the song actually went through that experience. And so I think that there's something about changing the tools, changing the paintbrush.
[0:39:55] But the fundamentals of how people connect and what people are excited by and why people spend their time the way that they do, they get to do more of it. And it can actually enhance the value in ways that were not present before. Nice. I should insert a quick story here. It's not my story. It's a story of thanks and also a story I learned about at lunch today. So I have to offer my deep thanks to Michael Chambers, who is a bioscience company founder down in Fargo, who made this whole
[0:40:23] So thanks to Michael. How about a round of applause for him? Even more impressive than Michael's contributions was the conversation we had with his eighth grade daughter, Eve, over lunch. And it got, Eve, it was this point, right? It was, she did her science experiment this year on how humans or whether humans can differentiate between artificially intelligently created, is that even a?
[0:40:51] Two adverbs tied together. Got to figure out a term. Created, whether it was poetry or I think art or food recipes versus human created. And also looking at the psychology as well. So it ties to that conversation really nicely too about what's distinctively human. Yep. Now, I do want to say that it's not all roses, right? I think that one thing about AI that I think is...
[0:41:18] I don't know whether every technology feels this way, but I really feel it for AI is that you cannot talk about the benefits without talking about the risks. You cannot talk about all the wonderful things that we'll do without also thinking about
[0:41:29] But what about in the wrong hands? What about if everyone's doing this thing in this particular way? And so I think the truth is that we as a society, again, need to pick. We need to think about how we want to navigate through this. And there's going to be ups and downs. There's going to be things that will feel like, oh, it used to be so great, and now it's just totally different. And I think that that balance, that is part of life. That is part of what
[0:41:54] Humans have been doing for thousands of years. And it's just so compressed in terms of AI, the progress and the rate at which we're developing it, and I think that the impact that we can have. And so I think that it is, again, just really important that everyone sort of engages with this. And so one of the best ways to answer this question of how to make sure it deepens the value of art is to actually do more AI art.
[0:42:17] And I think that with that kind of education and that kind of experience and do the science projects, I think that that is how we can get to a great world. Nice. And Andrew, an environmental management major from Penn Valley, California, continues this theme on kind of the humanity. And this is kind of a how question, but what do you think the difference is between human creativity and the way that ChatGPT creates text? Yeah, I suspect that they're less different than you think.
[0:42:46] And once again, I think it's hard to be totally definitive here. One thing that's interesting about deep learning and neural networks is that they were originally designed, neural networks were designed in 1942 or something like that, to be a model of the information processing of the brain. And throughout the history of them, people have said, the biological analogy, it doesn't hold. There's just no way.
[0:43:11] This thing is totally different from these artificial neural nets. They don't capture so much of the temporal information. There's just so many details that are different about the neural nets. And yet, here we are about 70, 80 years later, and actually you're basically using the same kinds of neural nets. We kind of made a little bit of improvements in some of the details, but fundamentally, if you were to show this to Pitts and McCullough, who invented this stuff in the 1940s, they'd be like, oh yeah, that was basically what I had in mind. And we're able to actually capture language. We haven't fully solved it, but we've really captured an important part.
[0:43:40] We're doing the same thing with images. We're doing the same thing with, you know, you ask Dolly for an image, it comes up with something totally new. And people like to say, oh, it's just next word prediction, right? The way we train these is that we show them text and we say, what comes next? What that misses is the way we actually train them is we show the AI text that's never seen before. And we say, well, what should come next? And this thing you've never seen. And if you think about it, you can phrase any super hard problem that way. You know, what is seven plus, so let's say you're given the sentence, the answer to seven plus five is?
[0:44:10] Blank. You've got to actually solve the problem. And you can imagine replacing that with any super hard problem. Or a really great idea for this character in this play, give the whole play, would be if then they went to blank. And you have to come up with something creative there. And so I think that this understanding, this sort of being able to predict what comes next.
[0:44:29] That's, I think, a fundamental thing. I think humans do it too. There's lots of research about how humans are prediction engines, right? That a lot of what you do, you think about when you're just kind of zoning out, and someone's talking, and you're thinking about what they're going to say next. You kind of know, and they say something surprising. You're like, wait, what? And I think that's going on with our neural nets too.
[0:44:47] And so I think that the mechanisms are way more coupled than you think. And the fact that these AIs, they are coming up with new things that have never been said before, never been thought before necessarily. And not that different, right? It's like an epsilon from all the things that have been thought and said. But I think that's true for humans too. Nice. And Jennifer, an English PhD student from Massachusetts, has our final student question. And this is about your work. What roles do creativity and hope
[0:45:15] Play in your research. Well, they're deeply, deeply intertwined, I'd say. I think hope, like the thing about research, right, is most of the time research fails. And you gotta be okay with that.
[0:45:27] And so you cannot do research successfully, as far as I know, at least the way that I know how to do it, without always having that hope, that optimism of you think about what if this works. And the fact of the whole company is fundamentally, you know, in some ways you can look at it as a, you know, it's very easy to think of the doomers and thinking about the, well, you know, all the bad things that could come from AI. But we are fundamentally a hope-based company, right? That we say...
[0:45:48] We want to unlock the benefits. We want to deliver them to everyone. We think that is possible. We think it can be done. We don't think it's easy. We don't think it will magically happen. We think it can happen through the efforts of us and everyone.
[0:46:01] That is such a deep part of our ethos of the day-to-day, the creativity of just coming up with new ways of approaching these problems that no one has done before and trying to instill that same thing into the AI. So that's what we're here for. Nice, nice. And very kind of great human answer to that tough question because I know with all the press going on with, hey, we have these tools, that sense of hope and creativity sometimes gets lost, right? So I would like...
[0:46:29] To turn the final comments over to you for anything, I probably missed something in my questioning, certainly. But there might be a message that you have for the members of the room here, in particular perhaps even the students. And it would be a great chance for you to close our discussion with your final thoughts. Absolutely. I think that maybe the most important thing to take away is that we are
[0:46:55] We're definitely going to go through a decade of change. And by the way, I don't think that it will stop. I think that AI will continue to progress. It will continue to change its character and to change our lives. But I think this next decade is kind of our first introduction. ChatGPT was the very first introduction for many people to this kind of technology and what it can do, and also what it can't do yet. And one thing that people get wrong is they look at what it can't do yet and say, ah, this is what it can't do. But that last word is very important.
[0:47:23] Because often, it's just like, yep, that's the next thing. The things that you kind of notice are the things that actually are right on the cusp of working. And so one of the most important things to think about how to navigate through this, both to help society, but also yourself, to think about how can I contribute, or how do I want to sort of be part of this world, and not just the AI world, but how to actually be as you enter society and get a job and do those kinds of things.
[0:47:51] Is to try to become an expert in this technology and just using it and understanding the ways to get the best out of it. And by the way, that will change. There's some skills that will directly transfer of, oh, like learning how to prompt ChachiBT in this specific way. There's lots of sort of magic that people do to get the best out of DALI where you just say, oh, it's like digital photo art or whatever. The specific magic words, the incantations.
[0:48:14] Those might change, right? But the underlying picture of here's an AI, which is this magical, mysterious thing in a lot of ways, and then just getting excited about it, thinking about your use case, what you're excited about doing, whether it's homework related, whether it's something in your community, and then really doing it and just spending a lot of time just sort of trying to figure out how to get the best out of it. And that's something I've noticed, that the people, even as the generations of these AIs have come,
[0:48:40] There seem to be a consistent set of people who just get the best out of them and do magical things. And so I think that that is the promise of the world that we're signed up for is I think everyone can become a magician and just, you know, it's like we're all sort of signed up for Harry Potter here and the AI is our wand. So hopefully that's a good analogy, but I love it. Yeah. Well, Greg, let me be the first to thank you for spending this time with us. A room full of 550 people who were, I could see, sitting on the edge of their seats about,
[0:49:09] Your view and your experiences and what the hope is for the future of this world living alongside as human beings with AI and the ever-present AI that will be here, as you say, from now until forever. So thank you so much. It's great to have...
[0:49:27] You returned to your first college campus, which is great. We claim you as ours. And just the experiences I know laid a great foundation for your future success. Undoubtedly, the support of your parents and the great wisdom and support that they've given you and your siblings and your wife, Anna, as well. So thanks to all of you for being here. And then in addition, we do have
[0:49:50] I showed it to you before, but a gift of UND swag. So you can wear it as a proud alumnus of the university, honorary alum. And then I just realized we also have these fancy UND Rubik's cubes. So you can take one of those as well, unlocking the mysteries. So how about a round of applause, a warm UND thanks. Greg Brockman. Greg, thank you. Thank you. Really appreciate it. Well done.