My guest today is Sam Altman. He of course is the CEO of OpenAI. He's been an entrepreneur and a leader in the tech industry for a long time, including running Y Combinator that did amazing things like funding Reddit, Dropbox, Airbnb. A little while after I recorded this episode, I was completely taken by surprise when at least briefly he was let go as the CEO of OpenAI. A lot happened in the days after the firing, including a show of support from nearly all of OpenAI's employees and Sam is back.
So before you hear the conversation that we had, let's check in with Sam and see how he's doing. Whoop, whoop, whoop. Hey, Sam. How are you? Oh man, it's been so crazy. I'm all right. It's a very exciting time. How's the team doing? I think a lot of people have remarked on the fact that the team has never felt more productive or more of a mistake or better. So I guess that's like the silver lining of all of this. In some sense, this was like a real moment of growing up for us. We are very motivated to become a better and sort of to become a company ready for the challenges in front of us.
Fantastic. So we won't be discussing that situation and the conversation. However, you will hear about Sam's commitment to build a safe and responsible AI. I hope you enjoy the conversation. Welcome to Unconfused Me, I'm Bill Gates. Today we're gonna focus mostly on AI because it's such an exciting thing and people are also concerned. Welcome, Sam. Thank you so much for having me.
You know, I was privileged to see your work as it evolved and I was very skeptical, like I did not expect JPT to get so good. And I'm still, it blows my mind and we don't really understand the encoding that we know the numbers, we can watch it multiply, but the idea of where is Shakespearean encoded? Do you think we'll gain an understanding of the representation? Oh, 100%. Okay, you know, trying to do this in a human brain is very hard. You could say it's a similar problem which is there, these neurons, they're connected, the connections are like moving and we're not gonna like slice up your brain and watch how it's evolving, but this we can perfectly x-ray and there has been some very good work on interpretability and I think there will be more over time.
So yeah, I think we will be able to understand these networks, but our current understanding is low. The little bits we do understand have, as you'd expect, been very helpful in improving these things. So we're all motivated to really understand the scientific curiosity aside, but the scale of these is so fast and it is, you know, we could also say where in your brain is Shakespeare encoded and how does that represent it? Yeah, we don't know. We don't really know. But it somehow feels like even less satisfying to say we don't know yet in these like masses of numbers that we're supposed to be able to perfectly x-ray and watch and do any tests we want on.
Yeah, I'm pretty sure within the next five years we'll understand it. And in terms of both training efficiency and accuracy, that understanding would let us do far better than we're able to do today, 100%. You know, you see this in a lot of the history of technology where someone makes an empirical discovery. They have no idea what's going on, but it clearly works. And then as the scientific understanding deepens, they can make it so much better. Yeah, no, that in physics, biology, it's sometimes just messing around and it's like, whoa, how does this actually come together?
In our case, we had, you know, someone that was, the guy that built GPT-1 sort of did it off by himself and saw this and it was somewhat impressive. But, you know, no deep understanding of how it worked or why it worked. And then it was, we got these scaling laws where we could predict how much better it was going to be. That was why when we told you we could do that demo, we were pretty confident it was gonna work. We hadn't trained the model, but we were pretty confident.
And that has led us to a bunch of attempts and better and better scientific understanding of what's going on, but it really came from a place of empirical result first. You know, when you look at the next two years, what do you think some of the key milestones will be? Multi-modality will definitely be important. We started- Which means speech and speech out. Speech and speech out images, eventually video, clearly people really want that. We launched images and audio and it had a much stronger response than we expected.
We'll be able to push that much further, but maybe the most important areas of progress will be around reason and ability. Right now, GPT-4 can reason in only extremely limited ways and also reliability. You know, if you ask GPT-4 most questions 10,000 times, one of those 10,000 is probably pretty good, but it doesn't always know which one. And you'd like to get the best response of 10,000 each time. So that'll be, that increase in reliability will be important. Customizability and personalization will also be very important. People want very different things out of GPT-4, different styles, different sets of assumptions will make all that possible. And then also, the ability to have it use your own data. So the ability to know about you, your email, your calendar, how you like appointments booked, connected to other outside data sources, all of that. Those will be some of the most important areas of improvement.
In the basic algorithm right now, just feed forward, multiply, so to generate every new word, it's essentially doing the same thing. I'll be interested if you ever get to the point where, you know, like solving a complex math equation where you might have to, you know, apply transformations and arbitrary number of times, that the control logic for the reasoning may have to be quite a bit more complex than just what we do today. At a minimum, it seems like we need some sort of adaptive compute. Right now, we spend, you know, the same amount of compute on each token, the dumb one, or like figuring out some complicated math. Yeah, when we say do the Riemann hypothesis, that deserves a lot of compute. The same compute as saying the. Right.
So at a minimum, we've got to get that to work. We may need much more sophisticated things beyond it. You and I were both part of a Senate education session. And I was pleased that I think about 30 senators came to that and, you know, helping them get up to speed, you know, since it's such a big change age. And I don't think, you know, we could ever say we did too much to draw the politicians in. And yet, you know, when they say, oh, you know, we blew it on social media, you know, we should do better. You know, social media, we still, that is an outstanding challenge that there are very negative elements to that in terms of polarization. And, you know, even now, I'm not sure how we deal with that.
I don't understand why the government was not able to be more effective around social media, but it seems worth trying to understand as a case study for what they're going to go through now with AI. No, it's a good, good case study. And when you talk about the regulation, is it clear to you what sort of regulations would be constructed? I think we're starting to figure that out. It would be very easy to put way too much regulation on this space. And, you know, you can look at lots of examples of where that's happened before. But also, if we are right, and we may turn out not to be, but if we are right, and this technology goes as far as we think it's going to go, it will impact society, geopolitical balance of power, so many things that for these still hypothetical, but future, extraordinarily powerful systems, not like GPT-4, but something with 100,000 or a million times the compute power of that, we have been socializing the idea of a global regulatory body that looks at those super powerful systems, because they do have such global impact.
And one model we talk about is something like the IAEA. So for nuclear energy, we decided the same thing. This needs a global agency of some sort because of the potential for global impact. I think that could make sense. There'll be a lot of shorter term issues of what are these models allowed to say and not say, how do we think about copyright? Different countries are going to think about those differently and that's fine. You know, some people think, okay, if there are models that are so powerful, we're scared of them, you know, the reason nuclear regulation works globally is basically everyone, at least on the civilian side, you know, wants to share safety practices and it has been fantastic. When you get over into the weapon side of nuclear, you know, you don't have that same thing. And so if the key is to stop the entire world from doing something dangerous, you'd almost want global government, which, you know, today for many issues like climate, terrorism, you know, we see that it's hard for us to cooperate and people even invoke sort of US-China competition to say why any notion of slowing down would be inappropriate.
So isn't it gonna be hard to, any idea of slowing down or going slow enough to be careful will be hard to enforce? Yeah, I think if it comes across as asking for a slowdown, that'll be really hard. If it is instead says, okay, any, do what you want, but any compute cluster above a certain extremely high power threshold and given the cost here, we're talking maybe five in the world, something like that, any cluster like that has got to submit to the equivalent of international weapons inspectors and the model there has to be made available for safety audit, pass some tests during training and before deployment. That feels possible to me. I wasn't that sure before, but I did a big trip around the world this year, talked ahead of state in many of the countries that would need to participate in this and there was almost universal support for it. So I think that's not gonna save us from everything. There are still gonna be things that are gonna go wrong with much smaller scale systems. In some cases, probably pretty badly wrong, but I think that can help us with the biggest tier of risks. I do think AI in the best case can help us with some hard for sure, hard problems, including polarization because potentially that breaks democracy and that would be a super bad thing. Right now, I guess we're looking a lot of productivity improvement from AI, which that's overwhelmingly a very good thing. Which areas are you most excited about?
Yeah, so first of all, I always think it's worth remembering that we're just sort of on this long continuous curve. So like right now we have AI systems that can do tasks. They certainly can't do jobs, but they can do tasks and there's productivity gain there. Eventually they'll be able to do more things that we think of like a job today. And we'll of course find new jobs and better jobs. And I totally believe that if you give people way more powerful tools, it's not just they can work a little faster, they can do qualitatively different things. And so right now maybe we can speed up a programmer 3x. It's about what we see. I mean, that's one of the categories that we're most excited about. It's working super well.
But if you make a programmer three times more effective, it's not just that they can write, they can do three times more stuff. It's that they can, at that high level of abstraction, using more of their brain power, they can now think of totally different things. And it's like, you know, going from punch cards to higher level languages didn't just let us program a little faster, let us do these qualitatively new things. And we're really seeing that. And so as we look at these next steps of things that can do a more complete task, you can like imagine a little agent that you can say, go write this whole program for me. I'll ask you a few questions along the way, but it won't just be writing that few functions at a time. That'll enable a bunch of new stuff.
And then again, it'll do even more complex stuff. Someday, maybe there's an AI where you can say, you know, go start and run this company for me. And then someday there's maybe an AI where you can say, like, go discover new physics. And it's the stuff that we're seeing now is very exciting and wonderful. But I think it's worth always put in context of this technology that, at least for the next five or 10 years will be on a very steep improvement curve. These are the stupidest the models will ever be. But coding is probably the area, the single area, from a productivity gain we're most excited about today. Massively deployed and, you know, at scale usage at this point, healthcare and education are two things that are coming up that curve that we're very excited about too. But the thing that is a little daunting is, unlike previous technology improvements, this one could improve very rapidly. And there's kind of no upper bound. I mean, the idea that it achieves human levels on a lot of areas of work, you know, even if it's not doing unique science, it, you know, can do support calls and sales calls.
I guess, you and I do have some concern, along with this good thing, that it'll force us to adapt faster than we've had to ever before. That's the scary part. It's not that we have to adapt. It's not that humanity is not super adaptable. We've been through these massive technological shifts and a massive percentage of the jobs that people do can change over a couple of generations. And over a couple of generations, we seem to absorb that just fine. We've seen that with the great technological revolutions of the past. Each technological revolution has gotten faster and this will be the fastest by far. And that's the part that I find potentially a little scary, is just the speed with which society is going to have to adapt and that the labor market will change. One aspect of AI is robotics or blue collar jobs when you get sort of hands and feet that are at human level capability. And, you know, the incredible chat GPT breakthrough has kind of gotten us focused on the white collar thing, which is super appropriate.
But I do worry people are losing the focus on the blue collar piece. So how do you see robotics? Super excited for that. We started robots too early. And so we had to put that project on hold. It was hard for the wrong reasons. It wasn't helping us make progress with the difficult parts of the ML research. And, you know, we were like dealing with bad simulators and breaking tendons and things like that. And also we realized more and more over time that what we really first needed was intelligence and cognition. And then we could figure out how to adapt it to physicality. And it was easier to start with that the way we've built these language models. But we have always planned to come back to it. We've started investing a little bit in robotics companies.
I think on the physical hardware side, there's finally, for the first time that I've ever seen, really exciting new platforms being built there. And at some point, we will be able to use our models as you are saying with their language understanding and future video understanding to say, all right, like, let's do amazing things with a robot. But if the hardware guys, you know, who've done a good job on legs, actually get the arms, hands, fingers, piece, and then we couple it, you know, and it's not ridiculously expensive, that could change the job market for a lot of the blue color type work pretty rapidly.
Certainly the prediction, like the consensus prediction, if we rewind seven or 10 years, was that the impact was going to be blue color work first, white color work second, creativity, maybe never, but certainly last, because that was magic and human. Obviously, it's gone exactly the direction. And I think there's like a lot of interesting takeaways about why that happened. You know, creative work actually, the hallucinations of the GPT models is a feature, not a bug, it lets you discover some new things. Whereas if you're, you know, having a robot move, having machinery around, you better be really precise with that.
And I think this is just a case of, you've got to follow where technology goes and you have preconceptions, but sometimes the science doesn't want to go that way. So what applications on your phone do you use the most? Slack. Really? Yeah. I wish I could say chat GPT. No, it's okay. Even more than like email. Way more than email. The only thing that I was thinking possibly was messages, but like I messages, but yeah, more than that. And so like inside opening eye, there's a lot of coordination going on. What about you? It's Outlook. I'm, you know, this old style email guy, either out of the browser, because of course a lot of my news is coming through the browser.
I didn't quite count the browsers in app. It's not possible I use it more, but I still don't, I still have that Slack. I'm like, we're, I'm on Slack all day. Incredible. Well, we've got a turntable here and I have Sam, like I have for other guests to bring one of his favorite records. So what have we got? So I brought the new four seasons of Vivaldi recomposed by Max Richter. I like music with no words for working. And this, it had like the old comfort of Vivaldi and like pieces I knew really well, but enough new notes that it was just a totally different experience. And there's these like pieces of music that you like form these strong emotional attachments to because you were like listening them a lot in a key period of your, of your life. This was something that I listened to a lot lot while we were starting opening eye. I think it's like very beautiful music. It's like soaring and optimistic and just like perfect for me for working. And I thought the new version is just super great.
Is it performed by an orchestra? It is the Chineche orchestra. Oh, fantastic. Should I play it? Yeah, let's. OK. This is like the intro to the song we're going for. MUSIC Do you wear headphones? I do. And do your colleagues give you a hard time about listening to classical music? I don't even know what I listen to because I do care phones. But it's very hard for me to work in silence. Like I can do it, but it's not my natural song. Yeah, no, it's fascinating. Songs with words, I agree. I would find that distracting. But this is more of a moon type thing. Yeah, and I put it, I have it quiet. Like I can't listen to a lot of music either. But it's somehow just always what I've done.
No, it's fantastic. Thanks for bringing it. You know, now with AI, to me, if you do get to the incredible capability, you know, AGI, AGI, plus, I guess I, you know, there's three things I worry about. One is that a bad guy is in control of the system. And so if we have good guys who have equally powerful systems, that hopefully minimizes that problem. There's the chance of the system taking control. And for some reasons, I'm less concerned about that. I'm glad other people are. The one that sort of befuddles me is human purpose.
I get a lot of excitement that, hey, I'm good at working on malaria and malaria eradication and getting smart people and applying resources to that. When the machine says to me, Bill, go play pickleball. I've got malaria eradication. You're just a slow thinker. Then, you know, it is a philosophically confusing thing. And how you organize society, yes, we're going to improve education, but education to do what if you get to this extreme, which we still have a big uncertainty.
But for the first time, the chance that might come in the next 20 years is not zero. There's a lot of psychologically difficult parts of working on the technology, but this is for me the most difficult. Because I also. Yeah, you have a lot of satisfaction from that. And it's like, in some real sense, this might be like the last hard thing I ever do. Well, our minds are so organized around scarcity, scarcity of teachers and doctors and good ideas that partly I do wonder if a generation that grows up without that scarcity will find the philosophical notion of how to organize society and what to do. Maybe they'll come up with a solution and I'm afraid my mind is so shaped around scarcity, I mean, to have a hard time thinking of it.
That's what I tell myself to and what I truly believe, that although we are giving something up here in some sense, we are gonna have things that are smarter than us. If we can get into this world of post-scarcity, we will find new things to do. They'll feel very different. Maybe instead of solving malaria, you're deciding which galaxy you'd like and what you're gonna do with it. I'm confident we're never gonna run out of problems and we're never gonna run out of different ways to find fulfillment and do things for each other and sort of understand how we play our human games for other humans in this way that's gonna remain really important. It's gonna be different for sure, but I think the only way out is through, we just have to go do this thing. It's gonna happen. This is like now an unstoppable technological course.
The value is too great and I'm pretty confident, very confident, we'll make it work, but it does feel like it's gonna all be so different. The way to apply this to certain current problems, like getting kids a tutor and helping to motivate them or discover drugs for Alzheimer's. I think it's pretty clear how to do that. Whether AI can help us go to war less, be less polarized, you'd think it should drive intelligence and not being polarized kind of is common sense and not having more as common sense, but I do think that's a lot of people would be skeptical. So I'd love to have people working on the hardest human problems, like whether we get along with each other.
You know, I think that would be extremely positive if we thought the AI could contribute to humans getting along with each other. I believe that it will surprise us on the upside. The technology will surprise us with how much it can do. We've gotta find out and see, but I'm very optimistic and I agree with you, what a contribution would that be? In terms of equity, technology is often expensive, like a PC or internet connection and it takes time to come down in cost. I guess the cost of running these AI systems, it looks pretty good that the cost per evaluation is gonna come down a lot. It's come down, enormous amount already. GPT-3, which is the model we've had out the longest and the most time to optimize. In the three years, that's three and a little bit years that's been out.
We've been able to bring the cost down by, I think a factor of 40. So for three years time, that's a pretty good start. For 3.5, we've brought it down. I would bet close to 10 at this point, four is newer, so we haven't had as much time to bring the cost down there, but we will continue to bring the cost down. I think we are on the steepest curve of cost reduction at ever of any technology. I know way better than Moore's Law. It's not only that we figure out how to make the models more efficient, but also as we understand the research better, we can get more knowledge, we can get more ability into a smaller model.
So I think we are gonna drive the cost of intelligence down to so close to zero, that it will be just this a foreign after transformation for society. Right now, my basic model of the world is cost of intelligence, cost of energy. Those are the two biggest inputs to quality of life, particularly for poor people, but overall, if you can drive both of those way down at the same time, the amount of stuff you can have, the amount of improvement you can deliver for people, it's quite enormous, and we are on a curve, at least for intelligence, we will really, really deliver on that promise. But even at the current cost, which again, this is the highest it will ever be in much more than we want.
For 20 bucks a month, you get a lot of GPT-4 access and way more than 20 bucks worth of value. So I think we're already like, we've come down pretty far. And what about the competition? Is that kind of a fun thing that many people are working on this all at once? It's both like annoying and motivating and fun. I'm sure you've felt similarly, but it does push us to be better and do faster and do things faster. We're very confident in our approach. We have a lot of people that I think are skating to where the puck was and we're going to where the puck is going and feels all right.
I think people would be surprised at how small open AI is. How many employees do you have? About 500, so we're a little bigger than before. That's why. Okay. But by Google Microsoft, Apple stands. It's tiny. And we have to not only run the research lab, but now we have to run a real business and product, two products.
So that the scaling of all your capacities, including talking to everybody in the world and listening to all those constituencies, that's got to be fascinating for you right now. It's very fascinating. Is it mostly a young company or is it an older company than average? Okay. It's not a bunch of 24-year-old programmers. It's true, my perspective is warped because I'm in my 60s.
I see you and you're younger, but you're right. It's 40. You have a lot in 40s. 30s, 40s, 50s. Yeah, so it's not the early Apple Microsoft, which we were really kids. Yeah, it's not. And I've reflected on that. I think companies have gotten older in general. And I don't know quite what to make of that. I think it's like somehow a bad sign for society. But I tracked this at YC and the best founders have trended older over time. Yeah, that's fascinating.
And then in our case, it's a little bit older than the average, even still. No, you have got to learn a lot by your whole Y Combinator helping these companies. I guess that was good training. That was super helpful. That was super helpful. Yeah. Including seeing mistakes. Totally.
OpenAI did a lot of things that are very against the standard YC advice. We took 4 1 1 2 years to launch our first product. We started the company without any idea of what a product would be. We were not talking to users. And I still don't recommend that for most companies. But having learned the rules and seen them at YC made me feel like I understood when and how and why we could break them. And we really did things that were just so different than any other company I've seen.
The key was the talent that you assembled and letting them be focused on the big, big problem, not some near term revenue thing. I think Silicon Valley investors would not have supported us at the level we needed. Because we had to spend so much capital on the research before getting to the product.
We just said, eventually the model will be good enough that we know it's going to be valuable to people. But we were very grateful for the partnership with Microsoft because this kind of way ahead of revenue investing is not something the venture capital industry is good at. No, and then capital costs were reasonably significant, almost at the edge of what venture would ever be comfortable with. Maybe past. Yeah, maybe past.
And I give you a sought to incredible credit for thinking through how do you take this brilliant AI organization and couple it into the large software company. And it has been very, very synergistic. It's been wonderful, yeah. You really touched on it, though, and this was something I learned from my commentator.
We just said, we are going to get the best people in the world at this. We are going to make sure that we're all aligned at where we're going in this AGI mission. But beyond that, we're going to let people do their thing and we're going to realize it's going to go through some twists and turns and take a while. And we had a theory that turned out to be roughly right, but a lot of the tactics along the way turned out to be super wrong.
And we just tried to follow the science. Yeah, I remember going and seeing the demonstration and thinking, OK, what's the path to revenue on that one? What does that like? And in these frenzied times, you're still holding on to an incredible team. Yeah, great people really want to work with great colleagues. And so there's a deep center of gravity there. And then also people, I mean, this sounds so cliche and every company says it, but people feel the mission so deeply.
Like everyone wants to be in the room for the creation of AGI. No, it must be exciting. And I can see the energy when you come up and blow me away again with the demos. I'm seeing new people, new ideas. You're continuing to move at a really incredible speed. What's the piece of advice you give most often? Well, I think there's so many different forms of talent. And really in my career, I thought, OK, just pure IQ, like engineering IQ. And of course, you can apply that to financial and sales.
But that turned out to be so wrong. And building teams where you have the right mix of skills is so important. And so getting people to think for their problem, how do they build that team that has all the different skills? That's probably the one that I think is the most helpful. I mean, yes, telling kids math, science is cool. If you like it. But it's that talent mix that really surprised me. What about you? What advice do you give? I think it's something about how people, most people are sort of miscalibrated on risk.
And they're afraid to leave the soft, cushy job behind to go do the thing they really want to do when, in fact, if they don't do that, they look back at them at their lives, like, man, I never went to go start this company when I'm a starter. I never tried to go be an AI researcher. I think that's sort of much riskier. And related to that, being clear about what you want to do and asking people for what you want, I think goes a surprising in the wrong way. And so I think a lot of people get trapped in spending their time and not the way they want to do.
And probably the most frequent advice I give is to try to fix that some way or other. Yeah, if you can get people to end a job where they feel they have a purpose, it's more fun. And sometimes that's how they can have gigantic impact. That's for sure. Well, thanks for coming. It was a fantastic conversation. And in the years ahead, I'm sure we'll get to talk a lot more as we try to shape the AI in the best way possible. Thanks a lot for having me. I really enjoyed it.
Unconfused Me is a production of the Gates Notes. Special thanks to my guest today, Sam Altman. Can you remind me what your first computer was? A Mac LCSU. Oh, nice choice. It was a good one. I still have it. Still works.