[0:00:04] Thank you both for being here. Neither of them probably needed an introduction. But of course, for the record, Sally Kornbluh became MIT's 18th president January 1, 2023. She's a cell biologist whose eight-year tenure at Duke's University as provost, earned her reputation as a brilliant administrator, a creative problem solver, a leading advocate for faculty excellence and student well-being.
[0:00:30] And in her first year of MIT, amongst many initiatives, she has focused as well on AI and has testified MIT's efforts to make sure that AI is broadly beneficial for society. Sam Altman is an entrepreneur, an investor, a programmer. He co-founded OpenAI in 2015. He is their CEO. And OpenAI is an AI research and deployment company whose mission is to ensure that artificial general intelligence benefits all of humanity.
[0:00:59] Sam and Sally, thank you for being here. I'm going to turn it over to both of you. Great. Thanks so much, Mark. Welcome. I will say I think this is the most popular event I've seen since I arrived at MIT. And I should say that the questions I'm asking were submitted by students and sort of gone through and curated. But this really reflects the curiosity of our community here and enthusiasm for you being here. So let me start with, oh, sorry. Thank you for having me.
[0:01:34] Absolutely. So let me just dive in. So according to columnist Kevin Ruse and others, in some circles, the question is, what is your PDoom? It's a common icebreaker, or so I'm told. Most of you probably know that PDoom, or probability of doom, is calculated on a scale from 1 to 100. And the higher the score you give, the more strongly you believe we'll end up in a doomsday scenario where AI eliminates all human life. So Sam, just to break the ice, what is your PDoom?
[0:02:04] I think it's sort of the badly formed question. Hey, I'm glad I didn't take any responsibility. I think it's a great question for people. It's a great way to sound smart and important. I have as much fun pontificating on numbers as anybody else, but whether you say it's 2 or 10 or 20 or 90, the point is it's not zero. And I think...
[0:02:35] Another reason I think it's a badly formed question is that it sort of assumes that it's a static system. I think what we need to do is find a way to make the future great, make the future exist, like not tolerate any branches of it. We would just sort of have the doom come to play. But I think society always holds space for doomsayers. There's value to that.
[0:03:03] I'm happy that they exist. I think it makes us think harder about what we're doing. But I think the better question is, what needs to happen to navigate safety sufficiently well? Right, so be aware as you're developing things of that possibility, but not indulge in them too much. More than aware, like really confront it and take it extremely seriously. Yes, fair enough.
[0:03:25] So, you know, you've been in this business for a little while now. How have your views of AI changed over the past decade? Five or 10 years ago, did you expect AI GPT would become as powerful as it has? Well, I honestly still think it's not very good. I think we will make it very good, but we have a ton of work in front of us. I think 10 years ago, I probably had a more naive conception of AI as this, like...
[0:03:54] And now I think of it much more like any other technological revolution. Hopefully the biggest and the best and the most important and the greatest benefits. But we have a new tool in the tech tree of humanity.
[0:04:22] And people are using it to create amazing things. I think it will continue to get way more capable and way more autonomous over time. But even then, I think it's just going to integrate into society in an important and transformative way. But something that is somehow going to be, I think if AGI got built tomorrow, and you asked me what would happen the next day, 10 years ago, I would have said, can't really imagine it. It should be just this like,
[0:04:52] Absolute transformation, singularity, everything is different all at once. And I now think it won't really be like that at all. So you think in some extent you were optimistic, but you couldn't have placed yourself in the current moment in either way, and it's the same for... I don't know, like optimistic or pessimistic, I think I was just like somewhat wrong. I mean, there were ways in which it was too optimistic and too pessimistic at the same time. You know, if we make something that is like, you know...
[0:05:22] As smart as all of the super smart students here. That's a great accomplishment in some sense. There's already a lot of smart people in the world. So maybe things go faster. Maybe quality of life goes up. Maybe the economy cycles a little bit faster. But if the rate of scientific discovery becomes 10 times faster than it is today, I don't know how different that will feel to us living through it at that time. That's interesting. That's interesting.
[0:05:50] So just stepping back a little bit and looking at the current models, thinking about AI systems like ChatGPT, what do you think is necessary to remove bias from the systems? And can you offer an example of thinking about bias that are in today's AI systems and how we might think about that going forward? I think we've made surprisingly good progress about how we can align a system to behave according to a certain set of values. I think...
[0:06:19] For as much as people love to talk about this and say that, oh, you can't use these things because they're just spewing toxic waste all the time, if you use GPT-4, it behaves kind of the way you want it to and reasonably well, and we're able to get it to follow, not perfectly well, but better than I at least thought was gonna be possible by this point, a given set of values. But that gets to a now harder question, which is who decides what
[0:06:47] What bias means and what values mean. How do we decide what the system is supposed to do? How much does society define broad bounds around the edges versus how much do we say, you as a user, we trust you to use the tool. Not everybody will use it in a way we like, but that's kind of the case of tools. I think it's important to give people a lot of control over how they use these tools.
[0:07:18] And even if that means that they may use them in ways that you or I don't always like. But there are some things that a system just shouldn't do. And we'll have to kind of collectively negotiate what those are. I mean, you know, it's interesting thinking about whether you can make the model sort of less biased than we are as human beings in a sense. Because, you know, you talk about things like, you know, in medicine, you know, bias against certain demographic groups, for instance.
[0:07:47] They're actually trained on the way our human doctors are behaving, right? They are, but then we do this RLHF step where we can exert quite a lot of influence. Humans are clearly very biased creatures and often unaware of it, and I don't think that GPT-4 or 5 shares our same psychological flaws. Probably it has its own different ones. But yeah, I think these systems can be way less biased than humans.
[0:08:17] Something to strive for. Aside from bias, other things that have sort of been part of the public consciousness in terms of concerns, et cetera, are privacy issues, which loom large for a lot of people when considering the future of LLMs. So how do we navigate the balance between personal privacy and the need for shared data to train AI models? I can imagine this future in which, if you want, you have a personalized AI.
[0:08:50] that has read every email, every text, every message you've ever sent or received, has an access to a full recording of your life, knows every document you've ever looked at, every TV show you've ever read, everything you've ever said or heard or seen, like all of your bits of input in and out. And you can imagine that that would be a super helpful thing to have. You can also imagine the privacy concerns that that would present. And I think
[0:09:19] If we stick on that frame and not say, well, should AI be able to train on this data or that data, but how are we going to navigate the privacy versus utility versus safety trade-offs or security trade-offs that come with that? And what does it even mean? Do we need a new definition of privileged information so that your AI companion never has to testify against you or can't be subpoenaed by a court? I don't even know what the problems are going to be.
[0:09:46] But this question of where we all will individually set the privacy versus utility trade-offs and the advantages that will be possible for someone to have if you say, I am going to let this thing train on my entire life. That's like a new thing for society to navigate. I don't know what the answers will be. I don't know where most people will make the trade-offs. I don't know what we'll say is even permissible in the bounds.
[0:10:12] We faced a little bit of that before about where we trade off some privacy for utility with the services we all use, but that can go so incredibly far with AI. There are all these things that we've had to negotiate with the internet, things about how we think about privacy, how we think about online ads, that when you intersect them with AI become much higher stakes and much bigger trade-offs that I think we're gonna start really facing. Yeah, no, that's interesting in terms of also
[0:10:42] How much individual control there's exerted? In other words, when you're talking about aggregated data, a good example when you talk about higher stakes, again, is sort of health record data, et cetera, and how much we can build into it some sort of personal ability to sort of set that sliding scale with how much information you're willing to have as part of that training. I think in that case, we are a little bit off in terms of how we have the conversation.
[0:11:10] What you want out of GPT-5 or 6 or whatever is for it to be the best reasoning engine possible. It is true that right now, the only way we currently know how to do that is by training on tons and tons of data. And in the process of that, it is learning something about how to do very, very limited reasoning or cognition or whatever you want to call it. But the fact that it can memorize
[0:11:37] Data or the fact that it's storing data at all in its parameter space, I think we'll look back and say that was kind of like a weird waste of resources. It is true that GPT-4 can kind of act like a database, but barely. It's slow, it's expensive, it doesn't work very well. It's not really what you want. It's just kind of as a side effect of the only way we know how to make a model, a reasoning engine right now, it has all these other properties.
[0:12:06] I assume at some point we'll figure out how to sort of separate the reasoning engine-ness from all of the needing tons of data or storing the data in there, and we'll be able to treat them as sort of separate things, and I think that'll make some of these privacy issues easier. No, that makes a lot of sense. Speaking about openness, there's been considerable discussion of whether OpenAI is actually open.
[0:12:32] You said that while it may not be completely open source, it's open in other ways. Can you say a little bit more about this and how you think about it? We make a great free AI tool available. Hundreds of millions of people, hopefully billions of people will use it in the future. We don't run ads. We just do this as like a public good because we think it's important to put the tool in people's hands. And we want it to be very widely available, very easy to use, very helpful. I think that's...
[0:13:01] Just a cool thing. It is a cool thing. You know, it is funny how it starts as you're talking about how AI gets built into us or our normal lives and that'll evolve. I mean, I think probably most of us remember the first time we saw ChatGPT and we're like, oh my God, that is so cool. Now we're trying to think about what the next generations, you know, what are the next generation of coolness going to be? By the way, I think that's great. I think it's awesome that for...
[0:13:35] Two weeks, everybody was freaking out about GPT-4 and thought it was super cool. And then by the third week, everyone was like, come on, where's GPT-5? I'm tired of waiting. I actually think that says something legitimately great about human expectation and striving and why we all have to continue to make things better. And I think it's great that a baby born today will never know a world in which the products and services they use are not intelligent, will never know a world in which cognition is not,
[0:14:05] Abundant and part of everything that you use. So I think this human discontent with the state of things and the expectation that the world should get better every year, I think that's awesome. Yeah, no, I agree.
[0:14:22] So in the past year, new electricity, speaking of less exciting edge of things, the new electricity demand from AI and data centers has been cited as an environmental concern. At the same time, many talk about AI assisting and decarbonization. What are your thoughts about this, the tension between its effect on climate, its ability to potentially help us fight the impact of climate change as it moves forward? I'll answer that specifically in a more general observation.
[0:14:54] It is true that AI needs a huge amount of energy, but not huge relative to what the rest of the world needs. If we have to spend, and I don't even think, if we spent 1% of the world's electricity training powerful AI, and that AI helped us figure out how to get to non-carbon based energy or do carbon capture better, that would be a massive win. And even if we didn't do that,
[0:15:22] If that 1% of compute that we spent on AI let people live their lives better and have to like... I read this thing about the compute Google used once compared to the amount of carbon that people used to spend driving in their cars places to get information. And you have people saying, Google's so horrible, we should shut it down. It's like spending all this energy. And it was intellectually a very dishonest thing to say because it was a net savings in energy, I think.
[0:15:51] Pretty clearly, the internet in general and the, you know, what it lets us do for telecommuting, probably also a savings. So, you know, for like AI, yeah, it's gonna need a lot of energy. We're gonna keep figuring out way more efficient algorithms, way more efficient chips. We're gonna get fusion. We're gonna power the stuff this way. So I think it is like important to address this issue, but we will in all of these fantastic ways.
[0:16:19] But I think this points to something else, which is the, you know, you opened asking about PDoom. And the level of Doomerism in society right now, I think the way we are teaching our young people that the world is totally screwed, that it's hopeless to try to solve problems, that all we can do is like sit in our bedrooms in the dark and think about how awful we are, is a really deeply unproductive streak. And I hope MIT is different than a lot of other college campuses. I assume it is, but you all need to like
[0:16:48] Make it part of your life mission to fight against this. Prosperity, abundance, a better life next year, a better life for our children. That is the only path forward. That is the only way to have a functioning society. And there will always be people who want to sit around and say, we shouldn't do AI because we may burn a little more carbon, or we shouldn't do AI because we haven't fully addressed bias. And it turns out a couple years later, we made a lot of progress on both of those things. And the anti-progress streak, the anti-people deserve a great life streak,
[0:17:17] Who are usually the people that have quite a lot of privilege in the first place is something I hope you all fight against. Yeah. You know, coming to MIT fairly fresh from the outside, I think this is sort of core to the MIT ethos, which is naming the problems and figuring out a way to solve them. I'm very happy to be here. Yeah, that's fantastic. We're very happy you are here. The other thing you said that really struck me as we think about the costs of AI, I don't just mean monetary costs, whether it's climate or anything else, it is this notion of deducting.
[0:17:52] What AI can contribute to the problem as sort of a long-term balancing of the ledger that I think is important. So that's interesting. Does OpenAI intend to build tools that will specifically impact science and engineering, or will you be more focused on sort of business and consumer applications? For sure, we intend to do that. I think the most, like personally, the thing I am most, most interested in
[0:18:22] It's how we use AI to increase the rate of scientific discovery. I believe that is the core engine of human progress and that it is the only way we drive the sustainable economic growth that we were talking about earlier. People aren't content with GPT-4. They want GPT-5. They want things to get better. Everyone wants more and better and faster. And science is how we get there. So of all of the things, of all the great things that AI will do, I am personally most passionate.
[0:18:51] about the impact that I hope it will have, expect it will have on science. That said, this may all be more one-dimensional than we think. If we make a great AI tool that can help people solve any kind of problem in front of them, that can help people reason in new ways, that's great for consumers, it's great for scientists, it's great for businesses, it's great for education. The G of AGI, the general, is sort of the surprising piece.
[0:19:21] Yeah, no, that's really interesting. And now I come back to your comment about, you know, sort of getting in your old car and driving to the library. You know, a lot of the creativity part is still human, but a lot of the aggregating all of the knowledge that you can use as a launching point can really be expedited, you know, by asking a few key questions to AI up front. Totally. I mean, again, what any one individual and certainly what any group of us will be capable of.
[0:19:51] I think it's going to, if we could go see what each of us can do 10 or 20 years in the future, I think it would astonish us today. It's like maybe in a few years, it's like each of us has a great chief of staff or a PhD student or whatever analogy you want that's off helping us optimize ourselves and do our best work and our best ideas and whatever. And then maybe at some point, it's like each of us has a full company.
[0:20:19] Full of like brilliant experts of anything. Just working super productively together. Cool. So what advice do you have for, we have a lot of
[0:20:30] Young researchers or people who are aspiring to be young researchers in the audience, what's sort of your general advice for making a real impact in the world? And you alluded to it in terms of thinking about possibilities and not sitting in your bedroom in the dark. I think that's a good base recommendation. But I'm just wondering what else you might want to say to this audience about that. First of all, I think this is probably the most exciting time to be launching your career.
[0:21:00] In many decades, maybe ever, I don't know, but it's like whatever it is, it's a really big deal. And the fact that you have this huge tailwind means I think you can, I think you take more risk than usual. I think if you do something that doesn't work out, there's just gonna be phenomenal opportunities for a long time. I think you can have more impact than normal, and so there's like a premium on having this be a period where you work really hard. I certainly would be biased to do something with AI, but.
[0:21:28] Of course I'm gonna say that, so maybe it's wrong. I think in general, the kind of core, the most important lesson to learn early on in your career is that you can kind of figure anything out. And that no one has all of the answers when they start out, but you just sort of stumble your way through it, have a fast iteration speed, try to drift towards the most interesting problems to you and be around.
[0:22:01] The most impressive people and have this like trust that you'll successively iterate to the right thing. And you can kind of like, you can do more than you think faster than you think. And people, it takes a while to learn that lesson, but it gets, you know, you see it work a few times and you really start to trust it. And so like, you can just do stuff.
[0:22:28] Sounds like not real advice or like very empty advice, but I think it's much more profound than it sounds on the surface. The other thing I would say is figuring out relatively early on, and this takes some practice, kind of like what your own personal, I don't even know what to call it, like passion, mission statement, like the kind of way you want to spend your time or what you really care about. And we talked about like this concept of like,
[0:22:58] Techno abundance as a way to drive prosperity and better lives for people, that's been something for me that has always really resonated and I've always tried to figure out how to work on that. But having some sort of, letting yourself develop some sort of guiding principle of how you make decisions about how to allocate your time and where to try to steer things, that's been very helpful to me.
[0:23:24] Yeah, I think this follow your passion and also, you know, from you're painting a picture of a world where there's sort of infinite possibilities and, you know, doing, you can't always be so strategic about what you think is good for you to do. You want to do something that imagines all those possibilities and follows those passions. Yeah, for me, it's like, like passion is not quite the right word. It's like something closer to like, what is the moral obligation for me to work on?
[0:23:55] And then on the really bad days when I'm not having fun, that somehow it's much more motivating than just the thing I like doing the most. That's interesting. Another element of this, and this is actually a really core part of MIT's culture, is entrepreneurship. And so we have a lot of aspiring entrepreneurs. So there's developing the sort of underlying ideas. But how do you think about building
[0:24:21] Successful companies in today's ecosystem and what part of the value chain where should new startups sort of focus their effort? Again, I think this is like the best time for new startups in particular in a very long time. Startups tend to succeed right around the time of big platform shifts. Big companies are slower and less innovative than startups, but they have a lot of other advantages.
[0:24:51] When you get the speed and iteration and cycle time advantage the most is when like the ground is shaking. And right now I think you can, there was like a moment like this right when the internet happened. There was a moment like this although smaller after mobile. There was another moment also smaller after AWS and this idea of like cloud services. And then for a very long time, like more than a decade, we've just been sort of waiting.
[0:25:19] And I think now we finally have a new platform. And so if history is a guide, which usually it is, and I suspect it will be this time, it's an amazing time to start a company. And the advantages you have as a company are you can move much faster. You can live in the future more than big companies that have quarterly or annual or whatever they have planning cycles. And that's how you win. And I think this is a great time to do it. Excellent. I think there's a lot of people here who,
[0:25:50] Take that to heart. Can I say one more thing about that? Please. I will issue like a, that was all the positive. Here's a warning. With any new tech platform, you can always drive phenomenal short-term growth. And so you have this class of AI startups like you used to have a class of mobile startups and before that like you used to have a class of internet startups that were not building an enduring business. But instead we're building this sort of like
[0:26:21] And you kind of delude yourself because you get amazing fast growth. And because there's this magic new technology and the dust hasn't settled yet, just because there's a magic new technology, it does not excuse you from the laws of physics of a business. You still have to figure out a way to build some sort of switching costs, some sort of relationship with customers, some sort of compounding.
[0:26:49] Yeah, that's really, that's I think important advice.
[0:27:09] The other thing, it's interesting how this question was phrased, and I'll read it in a minute, but now that I've heard you talk a little bit, I might phrase it a little bit differently. The question was phrased, in what ways might technology like ChatGPT threaten versus help the future of work? But it sounds like you tilt much more towards help, but also thinking about what that means in real terms. How does it help people in their future employment? One of the things that annoys me most about
[0:27:37] People who work on AI is when they stand up and with a straight face say, oh, this will never cause any job elimination. This is just an additive thing. It's all going to be great. This is going to eliminate a lot of current jobs. And this is going to change the way that a lot of current jobs function. And this is going to create entirely new jobs. That always happens with technology. It's probably never happened this fast. Although, again, we may be drinking the Kool-Aid too much.
[0:28:08] And the inertia of society may be such that it's slower than we think. But I kind of expect we're only a generation or two away from models that, for the first time, show some degree of real economic impact, good and bad, but something you can measure. And there will be classes of jobs that totally go away. There will be classes of jobs where you have to change what you do a lot. There will be classes of jobs where the productivity, compensation, whatever you want to talk about, whatever measure.
[0:28:38] It goes up by like a giant factor. And then there will be things that feel like jobs to the people of the future that to us today look like a complete indulgence and waste of time, as what many of us do today would look like to people from hundreds of years ago. I think as long as you believe that humans very deeply want to create and be useful and feel like they're making
[0:29:11] Like relative differential progress, all of which are things I would bet on hard. We're not going to run out of things to do. I love reading contemporaneous accounts from people living through previous technological revolutions and what they say about, man, we're all going to only work four hours a week if we work at all. And we're just going to, you know, like you say it every time. In some way, it does feel to me like this time is different. And as a matter of degree,
[0:29:43] It might be. And as a matter of speed, I really think it will be. And I have some concern about how quickly we can adapt to this kind of change. But I have no real concern that we can eventually adapt to this kind of change. I'm sure the social contract will change. I'm sure most jobs will be different in the future than they are today. But the deep human drivers don't seem to me likely to go anywhere.
[0:30:12] Interesting. And obviously, different categories of jobs are going to be really differentially affected. For sure. So with President Biden's recent executive order on AI, as well as the congressional hearings on AI regulation, there is a concern that regulatory frameworks might solidify the position of established players, might stifle innovation, competition, accessibility. How do you envision AI regulation, because we really are at a critical moment, being designed to uphold innovation and competition?
[0:30:40] While ensuring that the field remains accessible for emerging players to pioneer the next transformative technologies. I think we've faced versions of this with other kinds of regulation. You want to know that the food you buy in a grocery store is unlikely to make you sick. And we kind of all agree that regulation there is good. But you also want to be able to grow food in your backyard without having to go through a bunch of hoops. And you get to do that too.
[0:31:11] I think for AI systems, there will be some threshold above which we say, okay, this system presents a level of risk that we don't want to take without reasonable safety precautions. And then I think there will be a level of AI systems where we say, even though there's going to be misuse, we should open source these and let people use them and there should be no regulatory burden on companies developing them.
[0:31:43] Because we're willing to make the innovation and freedom trade-off for the negative safety consequences at level X. And then level Y can be totally different. And I totally get the impulse to say any regulatory action is unacceptable because it's just big companies that are going to use it for regulatory capture.
[0:32:10] And, you know, if society decides we don't want to regulate AI at all and we'll just take our chances, I'll accept the outcome of a democratic process. It seems to me good to have some voices saying, like, well, you open with a P. Doom question. Yeah, exactly. If the framing of that question is correct, then it seems to me useful to have some voices saying, let's not act out of fear, but proceed with some reasonable caution.
[0:32:40] Now that makes sense. Speaking of which, you've said you believe that the upcoming presidential election won't be the same as the last one. I think we all think that. But there are lessons to be learned from 2020. What are these lessons? How can we mitigate the risk AI poses to the democratic process and to the future of democracy in America? You know, maybe the use of advanced AI will be the least interesting thing about this election, the way it's shaping up.
[0:33:15] I think there will be better deepfakes, of course, and there will be better troll farms, of course. What I think is more interesting is trying to get ahead of, to the degree that we can, this is easier said than done, but trying to get ahead of the new things that just weren't possible before. So, like, customized one-on-one persuasion, where an AI system reads...
[0:33:47] All of your social media posts and target something just at you. That wasn't really possible with all of the sort of like online disinformation and trolling of the last election. And that's the kind of new thing that I wish we were taking more seriously. Yeah, I mean, that's the logical extension, as you mentioned, of advertisements coming up when you're surfing the web. People know you buy X kinds of shoes or whatever, how you think about things. Absolutely. A more local question.
[0:34:15] One of MIT's educational priorities is to train tomorrow's leaders to be, in essence, computing bilingual, meaning that regardless of their chosen field, they will need to be fluent in computer science and AI to advance their work. Can you comment on the impact of our way of thinking? I don't know if people talk to you about blended computing, thinking about how we train bilinguals, how they can start thinking about the future careers, really learning two different fields and using computing as a way into
[0:34:45] Other areas. One of the general observations I make about the history of computing, it has gotten increasingly more accessible and more natural over time. You know, like, I heard these stories of people with, like, punch cards that they would, like, have these crazy systems about how they sorted them and they would drop them and that was a big mess and it was, like, that was, like, not a natural thing to use.
[0:35:13] That's my undergraduate education. Sounds tough. Low-level programming languages were a step forward, but still not a thing that most of the world knew how to use, or I think a very natural tool in some sense. And then you get to current programming languages, and they're way more accessible and way easier to use, also way more expressive and more powerful. Along with that evolution, you go from the command line.
[0:35:42] to a GUI, to a mouse and a keyboard, to just touching your phone. That was pretty natural. And then you go to language, which is super natural. People are very good at language. As a way to use the computer, you can sort of ask ChatGPT something, but also a way to program the computer. So you saw these things converge to this one interface. And you don't have to be that bilingual anymore.
[0:36:12] You get the same way that you talk to a friend or a colleague. You can talk to a computer. And this is, I think, a more profound thing than it sounds like on the surface. The degree to which we can push AI and people to have the same kind of interface. So I'm more excited about humanoid robots than I am about other forms because I think the world is just very designed for humans and we should keep it that way. But we want the benefits of robots that can help us.
[0:36:42] I think we want AI systems to do their cognition and language, to communicate with us in language. It's a very human-focused thing. So my hope is we don't all have to be bilingual. It's interesting, I mean, because we have, I would say, an overwhelming number of our students are obviously interested in CS, and now that we've sort of rolled out these blended areas, I think it may be sort of the...
[0:37:10] Sort of first generation of what you're talking about. In other words, people who then, you know, next generations, it will be completely intuitive. Totally. Yeah, so that's really interesting. Let's see. There's some backup things here that other folks submitted. So how do you think AI will impact the financial sector? Thinking about sort of banking and equity. Have you thought about that at all? First of all, if we're...
[0:37:38] If we're on to the back of questions and we want to let people shout out, if you want to do that, I'm totally down. Otherwise, we can do them either way. Let me see. We can do some. Well, let me ask you that. Well, why don't you answer that? I have a couple other things. And then if people really feel compelled to yell out some questions related to this interview, then we can do that. But go ahead.
[0:38:04] I haven't thought as much as I would like to about any specific area because figuring out how to get the general purpose intelligence and what that means has been pretty all-consuming. You know, like education and healthcare are kind of maybe the two specifics that I've thought about the most. On something like the financial system, I expect AI to like impact that kind of as much as everything else.
[0:38:31] But I don't think I have a deeply insightful, specific thing to say about here's how this really transforms. Let's think a little bit more then about the educational sector. We think a lot about how people worry about things about AI in the classroom, but I really think there's just a huge potential for how we teach and how we tailor things to individuals, et cetera. So if you could say a word about that, I think that would be great. I think with what you're seeing people do already, just with like,
[0:39:03] Regular GPT-4, in-chat GPT, when you say like, please pretend you're a tutor and help me learn this thing. If that can work so well, then as people start to take next generation models and customize them for learning experiences, we're going to be in a very good place. And it's like pretty awesome to see what people are building already. It's great to hear from teachers about the impact it's had on their students. It's great to hear from students about what they're learning on their own.
[0:39:33] But this one seems like a kind of slam dunk good use. No, absolutely. So this isn't one of the sort of pre-submitted questions, but I'm always really interested when I meet people who are doing really interesting things, how you came to this. In other words, what was your sort of path? And how did you think about and make choices as you came to what you're doing? Well, I was like a very nerdy kid. I spent a lot of time reading sci-fi or watching Star Trek or whatever.
[0:40:05] That was already a time where all contemporary sci-fi was pretty dystopic. But the older stuff, the older Star Trek episodes, those were still kind of optimistic and cool. And you sort of saw how AI was going to be great and how that was obviously the future. And I kind of abstractedly thought, I would love to work on that someday. That would be so cool. I never thought it would actually happen.
[0:40:33] But I was always interested. And then life went on, I got a little bit older. I went to school and I decided to study in the AI lab. I also got really interested in energy, but this was kind of when I started to believe that if we could use technology to deliver abundance, that would not solve every problem. People would still find all sorts of ways to be unhappy, but it was something I really wanted to do.
[0:41:03] Worked in the AI lab. It was clear that AI was nowhere close to working. It was like back in 2004. Got sidetracked a bunch and went off and did a bunch of other stuff. Did a startup, became a startup investor. And then in 2012, noticed the AlexNet paper. It took me a couple of years to really internalize maybe this thing was finally working. And I was like, should really do something here.
[0:41:31] Excellent. Now, I think there's probably some question on the minds of some of the people here. What do you look for when you hire somebody? I mean, there'll be people sitting in this audience who are wondering that intensely. First of all, we would really love to hear from you. And I think this is probably the most or second most exciting time in OpenAI history right now. So this would be a good time to give us a call.
[0:41:57] Number up on the screen. What did we look for? Kind of all the obvious stuff, like original thinkers, smart, driven, dedicated people, people who sort of like are particularly driven to work on AI versus, you know, there's always like a set of people who will go work on whatever like the hot area is. And then there's like a set of people who are like, no, this is like the thing. And we like people for whom this is their thing. But that's kind of it.
[0:42:29] I mean, all of the obvious things and nothing that unusual. Interesting. If you were doing something else, what would it be? I would go work on Fusion full-time. Fusion. There's definitely people here who are interested in that as well. You know, we do have a few minutes. Yeah, so we don't have any mics. I see Manolis' hands up right away. But you've got to shout. I'll repeat the question. How do I...
[0:43:10] How do I want to plan my time for the next five years while humans are still helpful and useful? And it's striking because I see AI as an extraordinary tool and I am so looking forward to the future of work where I can do amazing things that I can't do right now, where all the mundane stuff that I'm doing now is replaced and I can do super awesome new things. But there will come a time
[0:43:37] Given the two curves of how fast human intelligence has progressed through the old school evolution and nurture, of course, and the curve with which AI is progressing, which is just ridiculous. So these curves are going to cross very soon. And I'm just curious, what are some unique capabilities that humans have that you think we will not be able to replicate with AI?
[0:44:03] What are some of the architectures that will help us maybe replicate those things and push beyond that? And how do you see the future of human versus machine intelligence more broadly? Two conflicting answers to that. Number one, I kind of suspect that forever onwards from now, it's always going to feel like, man, this next five-year period is so critical. This is when I can really contribute. And after that, who knows what happens. But in practice, there's always going to be
[0:44:33] It's always gonna feel like, man, this curve of intelligence is rising so fast. Right now I can use these tools to out-achieve, but eventually they outrun it. But at any point on that exponential, you'll always be able to use the tools to do amazing things, and it'll always feel like then you get totally outrun, but you never quite do because we just become more capable. And I didn't used to really think this way. I'm still not sure I'm right.
[0:45:02] But it does seem to me like we will just be able to accomplish more, do more things. We'll have ridiculous, the expectations will go up too to participate in the economy in some sense. But what we can get done before we feel like this wave is gonna crash over us, we'll always feel like, man, in this five years, this is my window, but it's gonna be a rolling five years forever. And I think we'll find that humans are really good at
[0:45:33] We're so wired to care about other humans. We're so wired to focus our energy, know what other people want, be focused on delivering value for others. I can see very extreme worlds where human money and machine money are just different things. But there's an increasing premium on the human money part of that world.
[0:46:04] I don't know exactly what it's going to look like, but I do kind of believe in the biological drivers of humanity not changing that much. And then the answer on the other side is, and this is just speaking personally, this is not a well-reasoned or logically defended thing at all. This is just what I feel. I do feel like I know I'm going to be nostalgic for this time. And it's sort of a strange thing to feel that while you're living through it. That's really interesting.
[0:46:37] I see another hand back there. Well, I'm going to have to ask you to just hand the microphone to whoever you see because I... Sure. More bullish on startups building end-user applications or infrastructure. Probably at this current phase of where we are on end-user applications, you've got to pick the right ones. You've got to pick the ones that benefit from the models getting better, not the ones that are betting the model actually doesn't get better and kind of are fixing the current generation problems. But I think it's like...
[0:47:32] I think there's a lot of value to unlock there. Building infrastructure can be great too. I think you can really succeed both ways, but the number of what feel like $100 billion application layer companies right now where you can really do something incredibly useful to people quickly seems pretty exciting to me. Yeah. Yeah.
[0:48:08] Thank you so much for giving this talk. I had a question in terms of the future that you see specifically for OpenAI. I was just wondering, you can come out with GPT like 6, 7, 8, 9, 10, 20, 30, 40, 50, 60. You have like your creation of Sora. A lot of companies are creating like life-size robots that are utilizing AI as the backend or a lot of technological companies also using ChatGPT as like a foundation for their own LLMs. So I was just wondering for you,
[0:48:36] Guys, specifically at OpenAI, what do you guys foresee as your niche in the future once you start, I guess, not perfecting, but getting to near perfection for your LLM models? I think it's so far, we're so far away from when we start to level off that that's not currently on our mind. I think at least for the next
[0:49:09] Three, four model generations. I believe we can make it so incredibly much better every time. We should focus on that. And if we can do that, everything else will kind of work out. Like I think our niche or whatever is we want to deliver great, useful, impressive cognitive capability for as abundantly and inexpensively as we can.
[0:49:34] And there's other good things we can do, but if we can just focus on that, I think it'll be like a great service to the world, and we can go on that for a while. I have two questions. One, where did you get your shoes? Adidas did a Lego collaboration. I love Legos. Amazing. Cool. My serious question is, how soon do you predict that we'll get to artificial general intelligence, and what is OpenAI's role in getting us there? You know, I don't... I don't...
[0:50:07] I no longer think there will be a time where the world agrees, okay, this was the year we crossed the AGI threshold. I think the phrase has become so overloaded as a definition. I think there are people who would say we'll get there soon. I think there's a lot of people who will say by 2040, when we have these unbelievably capable systems, it's not quite AGI yet, it can't do this one thing. So I think the question is,
[0:50:37] The only way I know how to form the question well at this point is what is the range of time that we get to capability X, Y, and Z. But I think that the AGI, I can't make myself do this because it's too deep in the OS at this point, but I try to not use the word AGI anymore. It's a total, I'm never going to succeed at that.
[0:51:03] It's like, you know, when can we do that new scientific discovery in some areas? When can we, like, add a lot of economic value? I don't know. I expect that, like, by the end of this decade, we have systems that create really significant economic value. Excellent. So maybe last question from the audience, or whoever's last one here, maybe. And then I'll close this out. Yeah. Thank you so much, Sam, for the talk.
[0:51:37] I guess I have a more general question. I guess what do you do when you have a problem to solve and you're stuck and you don't know the way forward? I guess what's your thought process on trying to get unstuck? I somehow try to change context. I try to talk to different people about it. I try to, in an extreme case, I'll go travel somewhere to really change things. The one time I like jet lag is if I'm really stuck on a problem and I wake up in the middle of the night in some new context and it seems to be helpful. But some way or other, I try to change context.
[0:52:08] Well, with that, I really want to thank you. That was fascinating. I think everybody greatly enjoyed the conversation. Thank you all very much. Thank you. Awesome. Can I sign this real quick?