So what version of you has died? Like, who were you when you started, and who have you become?
I was like much more carefree. Like when we started OpenAI, I was just like, you know, I had a very easy life at the time. I had a lot of hobbies, and you know, now I'm like incredibly precisely scheduled, and I have not gotten to do anything spontaneous in a long time.
Was there any point ever in your life where nobody else was convinced but just you?
Launching ChatGPT was kind of an example of that. Not that people didn't want to launch a chatbot, but most people thought it was too early and not good enough.
When you talk to other people, do they tell you that you're delusional sometimes?
Yes, but I think I should probably get more of that.
Well, one way that like a lot of people in roles like mine have had things go wrong is they get surrounded by too many yes-men.
You've gone through like people suing you, people dropping out, the government like trying to find out so many things about you. Have you changed as a person with all of this now?
There's no way that anyone was going to like get to build superintelligence and not have a crazy life, but I think I'm kind of like well-wired for it. Like I can deal with a lot of insanity whipping around me.
When I talk to somebody in the US, they're slightly pessimistic about AI. And when I talk to people in India, they're very optimistic about it.
It's an incredible difference. India is like our biggest market in the world.
Why do you think it's the difference?
One thing is that countries that are developing faster, where the economy is growing faster and there's like more of a need for products and services, are more welcoming of AI than more developed economies. So I'm sure that's part of it.
Do you think you're still being misunderstood today?
This technology is still not very well understood, but I think people at least understand AI is happening worldwide. This is like certainly the best time I've ever seen to make a new company.
Do you worry about it, OpenAI?
What's your biggest worry?
We have done something. We will continue to do something that is going to be an extraordinary force in society. I believe it can be a force for good, but you know, it's like very high stakes, and we have like a lot of responsibility.
Do you feel that when you're gone, you're not there to guide or protect your kids or family? Would you be very comfortable with your kids growing up in an AI-powered world which you're helping create today?
Before we start today's episode, I just have one request. Please subscribe to our channel. Me and my team, we work every day and try to bring the most valuable conversations from the world. All we want in return is just one subscribe from you.
On Figuring Out, we try to study how the world's most extraordinary minds think, and that's exactly what today's episode is about. Today's episode with Sam Altman is not about AI, is not about technology, it's about a human being who's building that technology. What are his insecurities? What are his fears? How does he think as a friend, as a father, as someone who's trying to build one of the most important things which will shape humanity? How does he notice patterns and find the non-obvious things to work on to build something incredible in the future? That's what we've tried to understand. I hope by the end of the podcast, you get to extract some lessons from his mind and implement in your life to make your life better. Watch this episode and let me know in the comments how did you find it. And the styling partner for this episode is Levi's. Click on the link in the description to explore their latest collection. Enjoy the show.
Welcome on the show, Sam Altman. Welcome on Figuring Out. It's an honor.
I really appreciate you doing this.
You know, if someone's watching you for the first time and they have no clue who you are, how would you explain to them who you are in the next 60 seconds?
I'm like a tech fan, like a curious person about the world, and a believer that AI is going to be this thing that touches all of our lives, hopefully in mostly super positive ways. I'm a dad. I really love my kids. That's what I would say.
But do you ever think about certain things that this is how you're going to think or say or speak about?
I do just kind of try to naturally talk about what's on my mind, or what I'm interested in.
Okay. And were you like this since you were a kid, always thinking about what's interesting to say next?
Because I think I was like that as a kid. I think I would just try to like, I think I just always like followed what I was curious about.
And as a kid, did you ever feel inferior to someone?
All the time. I mean, I'm sure most kids do. But I was like, you know, I was like a very nerdy kid. I like wasn't good at sports, and I grew up at a time when like that was very much the cool thing.
I was not good at sports either.
And it was like the cool thing.
Yeah. So yeah, total. I was like not a, I was not, I was, I always had like a lot of friends and I connected well with people, but I was never like a cool kid.
Okay. Did that bother you?
Sort of. I mean, it did not like, I don't think it bothered me like it bothers some people, but yeah, I think all kids like want to be popular and cool, and there's like, I think it's like very painful as a kid to not feel like in the inner circle or whatever.
You said it didn't bother you as it would to some people. What do you mean by that? Well, what's the distinction? What did you think about it?
I've talked to other people who are like, ah, you know, the defining characteristic of my childhood was that I was like a nerd and not cool, and I hated that. And I was like a nerd and not cool, and it like slightly bothered me, but I had my computer and I had like my little core set of friends, and I, you know, like I had a wonderful childhood, and I think I was like very happy.
Okay. So, here's what I want to do. I want to talk more about your brain, like how you think, how you actually make decisions, and what you think about certain things, because I think that's really fascinating to me.
And it's great underexplored.
I think it's really hard to like talk about yourself this way, but I'll try my best.
Yeah. Maybe just the kind of things you, because you've said a couple of things in like a lot of your interviews here and there, and I don't know why people didn't just go like deep down on that alley. Something I want to do. You know, when I talk to somebody in the US, a random person, they're slightly pessimistic about AI, and when I talk to people in India, they're very optimistic about it.
It's an incredible difference.
Why do you think it's the difference?
Well, first of all, India is like our biggest market in the world by total users and growing very quickly, and you know, there's incredible builder energy of people using it.
Yeah, more just India than the US. And I mean, people are just using it in incredible ways to build companies, or you know, I just heard about some Indian kid that needed to do some fundraising for his school and had never done fundraising before, and like was able to get like the full stack of it done with a set of our tools, and it was awesome. So I think it's like a country that has, there's just a lot of adoption. It's growing very quickly. But I do think that the US, at least on expressed preference, is unusually negative on AI relative to the rest of the world, and India is maybe unusually positive.
But why? Like, what do you make out of it? Why is it, why there's a difference?
One thing that many people have observed is that countries that are developing faster, where the economy is growing faster and there's like more of a need for products and services, and like you know, there's just huge undersupply, are more welcoming of AI than more developed economies. So I'm sure that's part of it.
Interesting. And the other thing, when I read about you, right, like I read about you in the US and I read about you in India, this sentiment of being pessimistic and optimistic, it also reflects about you, and that's the most fascinating thing, that you're this person who started this most remarkable technological revolution of our age. You were the first guy. You don't even own equity in your own company. And yet all of a sudden you became the bad guy. Like how? Why?
Well, I think the US has great skepticism of people in power, and in particular of like the just enormous power that the tech industry holds. But you know, I would say like this is probably a good thing in some ways. It does feel to me like it's gone a little far.
Interesting. What do you think is the biggest misunderstanding about you in the world?
Well, I mean, you touched on one of them, but like, I think not taking, not having equity in OpenAI has been a mistake. I did it for what I thought were good reasons. I think it's just been hugely confusing to people. I like, it's kind of like interesting to me, right? Is like if an artist works super hard on their art project or a sports star like practices super hard and doesn't make a ton of money from it, no one has like a problem with that. That's kind of understood, and this is like my equivalent of that. But yeah, like it's been...
Why is it misunderstood? What do people not like, what do they not get about it?
I think it's just very outside of the norm of why people assume that people start companies or, you know, work as business people.
It's just weird. And like people don't like weird.
It's weird to me as well.
Yeah. And then you're like, well, what, like why is he really doing it? Is it about, you know? Yeah. And why are you doing it?
I think this is the most interesting work and kind of the most important and useful thing I can imagine doing. It's incredibly interesting. It's incredibly exciting to get to do. Like, I mean, what a quest.
True. But it's a rabbit hole we'll just keep going into because I don't understand this at all, just like a lot of other people. But I really respect what you do. So thank you.
Okay. One of the things you said, you know, it takes like, like I'll read it exactly as it is: that you should trade being short-term low status for being long-term high status.
Yeah, which most people seem unwilling to do, right? So, as long as you're right, being misunderstood by most people is a strength, not weakness.
I find it fascinating. But the question is how, before the outcome is clear, how do you know that this is being misunderstood and this not being wrong?
Well, I mean, the hard part is you don't really know. You can have like a high degree of conviction. You can have an increasing degree of conviction as you get more evidence, which I think is the right thing to go for, but you don't know for sure. That's part of the excitement, like that you don't know for sure. But when we started, when we first started OpenAI, it was an extremely low status thing to do. Like we knew it'd be valuable if we get it to work, but almost no one took us seriously. It was like, okay, AGI, maybe 100 years from now, maybe never. You're a bunch of like kids, and you don't have the polished academics and whatever, and very, very negative. We had like a reason to make the bet, and as we figured more things out, we had a reason to be more confident, but it's still kind of amazing it all worked. So I think it was like a good expected value bet to take, but it was far, far, far from certain.
But what were the signals that was making you very certain that this might work? This something that I want to do.
Yeah. I kind of think all of us get like a few, a small number of like great insights over our lives that we apply again and again in new ways. And one of them for me was the sort of surprising properties of scale. I noticed this in companies, countries, cities, individual people's careers, lots of pieces of technology that you know, networks like Y Combinator, which was the thing I was doing before that, when you really scale up something that's working at small scale, you often see big super-linear returns to doing it at bigger scale. And because like anybody else, you know, I'm a pattern matcher whether I mean to do it or not. Once we saw that neural networks were working at all and they were getting better with scale, I was like very interested, and I had a strong intuition that we should try to scale them up a lot more. Now, this was before we knew how precisely we could predict properties with scale. This was years before we had the scaling laws. But we had this observation approximately that if you made a neural network bigger, it could do more impressive things, and this had first led to the AlexNet result and then many more things after.
But then these were the signals. Probably a lot of people must be seeing, right? What was preventing them to build on it?
I find this one of the most confusing things I have ever observed in my career. Why in 2015, 2016, and then even by 2018 to 2021, why other people were looking at the same data and not saying, this is crazy, like we have to invest much more in this. It seemed so obvious to me and everybody else in the building. I mean, maybe like almost we felt delusional because we're like, how can anybody not, how can people not see this? I think there was something about it almost felt too easy. It almost felt too unfathomable. You know, you'd look at the mistakes that these neural networks were making, and you would say there's no way that something so dumb can just fix all of its problems by scaling it up. It just, it seemed like all these smart people were saying you have to have symbolic representations or you have to do this or I can give you this very good reason. I remember someone articulating why our approach was never going to work, and it sounded so convincing.
It was like, it went something like, all right, what these models are doing is they are learning to predict from what's in their training set what's likely to come next. But that means they're never going to be able to make new knowledge because they can only model what's in their training set. And like that sounds very, you know, reasonable. Maybe you can show a model a million math problems and it can learn how to do those math problems, but you didn't show it a new thing that wasn't yet proven because it wasn't yet known, and so it wasn't going to be able to do that. And obviously now we know that's ridiculous. But at the time it sounded very compelling. And the thing that it was missing is that what these models are learning is not, I mean also is, but the important thing they're learning is not these specific math problems, but the general purpose ability of reasoning, and that can expand into new things.
So take me to a specific decision when you were doing all of this and other people were telling you that hey, this might never work out, right? Tell me to one, take me to one consequential decision that you took with 50% certainty and 50% delusion.
After we finished training GPT-4. This was about like, I think it was August of 2022.
And we finished it about seven or eight months before we released it to the world, something like that. But we were sitting on it and we were looking at it, and we knew it was extraordinary. But I don't, it took us a little while to realize just like the new avenues of research this was going to open up and how quickly the models were likely to improve post-GPT-4, and that we could do what became the chain of thought work and the reasoning models. And we then decided we were going to go make a very big bet on compute because we saw that we could continue to scale models and that the value of these models would be really important. But it was new, and the rest of the world wasn't there yet. They hadn't even seen GPT-4. They had not internalized what we thought it could do. And so we decided we would bet the company on buying up a lot of compute. And that felt very crazy to people. But we had enough conviction that we were probably right because we were able to use this model and we had internalized the scaling properties.
And what was the other option? You had a much less ambitious bet on compute and a slower rate of scaling.
And do you think if any amount of other information you would have gotten at that point, you would have changed the decision?
I think if the model were even a little bit less smart, you know, if GPT-4 had come out at like the GPT-3.8 scale, yeah, we would not have had the conviction to do such a big bet.
When you talk to other people, do they tell you that you're delusional sometimes even today?
Yes, but I think I should probably get more of that.
Well, I think one way that like a lot of people in roles like mine have had things go wrong is they get surrounded by too many yes-men.
And you really, you need the pushback. But how do you make sure that today you are not surrounded with yes-men?
Having your old friends around is very helpful. Having people that you've worked with for, you know, 10 plus years is very useful. And then really cultivating a culture where you reward people for pushing back hard when you're wrong and making that easy to do, I think is also important. But it takes like active care. You know, the more times that you've been right and the more sort of like power in the world you have, I think the harder this is to get.
See, here's the confusion as a founder. So when you have like all old friends, really smart people, the mentors that you think about, you have them around you and they're telling you that, hey, this is wrong and you're thinking in the wrong direction. How do you then convince them or probably just go against them? And how do you know that going against them is the right thing today? Because I'm sure this happens.
The number of times that I do something against everyone else's advice is very small. Like it happens, but it's rare. And I like, there are, I have met a lot of people, not a lot, I've met some people and founders in particular who take the most pride in doing something when everyone else tells them it's wrong. And my own belief is you should only do that very grudgingly. Like if you surround yourself by smart people, you should mostly listen to them. And if you decide that even though none of them think what you're doing is the right thing to do and you're going to do it anyway, you should like view that as an expensive, unlikely to be right kind of high stakes thing. And you really want to have conviction on that.
And talking about smart people and other founders, what do you think? What stops them? They're intelligent, clearly intelligent. What stops them to build such extraordinary things like you've done?
Well, I think AI has happened only once. So I think we got lucky or did a good job or whatever on problem selection. But I think a lot of people build extraordinary things. The number, the rate of great companies being created right now, I have never seen it be so high worldwide. I think this is like certainly the best time I've ever seen. I believe it is probably the best time in history to be able to build something, to make a new company, to make a new technology, to make a new kind of creative project. We have just gotten this unbelievable gift, and I think it's incredible what people are doing right now. So I think we're going to see a lot of very extraordinary things, companies and otherwise, get built.
So now that everyone's building companies and there are so many things happening, the action is happening, some founders watching this right now is thinking that things that are obvious to me because of AI and technology is not obvious to a lot of people. How can that person judge that it's the right thing to do and it's not being misunderstood?
I mean, this is kind of what you get paid to do as a founder, right? Like this is where if you are starting the most ambitious company you could start, you'll certainly be misunderstood.
Now, there's many other companies that are obviously good ideas, and you can still succeed big time at those. But if the ideas are obvious, a lot of other ambitious people are going to start them too. So when we started OpenAI in 2015, I think you could count on one hand the number of AGI labs that were being started that year. I mean, maybe you could count it on much less than one hand. Maybe it was only OpenAI that year. I'm not sure. And yet there were probably like 10,000 photo sharing companies. Most people thought starting a photo sharing company was a better idea. But the problem is there were 10,000 others, and also Instagram had a big network effect and whatever else. So I think picking things that you have a reason to believe are more valuable than they seem, even if they're unlikely to work, so valuable if they work that like the expected value is good, you have a real advantage when you find something like that. When you understand a technology or a piece of science or something about the world before other people do, there's some instinct which I understand to say, ah, you know, there's all these smart people, and I don't want them to call me dumb, so I, you know, they don't get it, and they're going to talk me out of it. But I think there's another instinct which I encourage people to cultivate, which is if you really understand something and it's not just like wishful thinking, but you can articulate why it's a good bet, you should cherish those moments. Like that's awesome as a builder. That's when you should be the most excited.
How do you decide that? How do you find a non-consensus bet?
Good question. I like, my own experience of this was this was one thing that just no one ever taught me. And you know, like very little of my experience in school was figuring out what questions to ask. You figure to like answer other people's questions, and then at some point I think like when you become a PhD student or whatever, you get to start asking more questions, and then eventually you do more and more of that. I think we should focus much more on teaching people, teaching very young people to figure out what questions to ask and what things are interesting to explore that we don't already know. And I think getting good at that skill was the best way that I figured out the sort of non-consensus bets to take.
Just figuring out how to ask good questions.
Figuring out how to ask good questions.
I think so, yeah. I mean, like all of these things, maybe not everybody is going to have equal aptitude for everything, but I think this is one that a lot of people seem to be able to learn. You know, there's, I forget which culture, which country has this rule where the kid, when he comes back from school, the mother asks the kid that what's the question that you asked today?
Do you know what culture? That's very interesting. You know what culture?
I think it's Israel or something. Someone in Israel. I'm not sure, but there is some culture where mothers just ask the kids, they don't ask about the highlight, they ask about what's one question did you ask today.
That's cool. Is that a good habit that we...
I'm going to start that with my kids. I'll report back in 15 years or something.
But how do you know that it's not wishful thinking? Is there a sign? Is there something that you can pin down to and be like, okay, if these are the signals in the thinking, the way I'm thinking, probably this makes sense.
Yeah. I mean, one thing is, can you explain it to an optimistic, enthusiastic, but kind of mentally disciplined smart friend, and can you update them about why it might be a good bet? So I remember when we were starting OpenAI, the default was like strongly skeptical, but I would talk to people and I would say, you know, here's the evidence, here's what we've seen, here's why this might work, here's why it'd be huge if it worked, and here's why this is an approach that we think has legs. And then, you know, if we can make superintelligence, it was kind of obvious why that would be valuable to make. And you know, not every time, but enough of the time when we talked to somebody who was like a smart person but a non-expert, they would say, 'Okay, that seems, you can, that updated me. That seems like it's worth doing.' Now, if we had convinced everybody, it probably would have been too obvious. But if we had not been able to convince anyone at all, I would have been more skeptical.
Interesting. Now, as I'm saying that out loud, I'm thinking of there must be historical examples of where someone was not able to convince anyone and they were still right. So at some point, you got to just like trust your own instincts here. Was there any point ever in your life where nobody else was convinced but just you?
I mean, there have been like small decisions where I was like, 'OpenAI is going to do this thing,' and none of the rest of the leadership team wanted to.
Launching ChatGPT was kind of an example of that. Not that people didn't want to launch a chatbot, but most people thought it was too early and not good enough.
Interesting. And what made you go against all of them and be like, 'Hey, let's just launch it'?
I think it was the YC ground. I think it was the YC experience. I think so. Like I had really, I had spent so long drilling it into founders that if you're not a little embarrassed of your first version, you're waiting too long to launch because what you need is feedback. The other thing that I had, because we had this like internal version, and this is actually a good insight. The retention was terrible of the internal version. I think it was like, you know, 5% or something of the first version of ChatGPT people would use and keep using. Now there's a bunch of reasons why it was bad data in the first place. We had better models, we had other interfaces. But many people would look at that and say you can't launch a product with 5% retention. That's terrible. And normally that's true. But then you would talk to those 5% of users, and some of them like really loved it. And there's this other YC learning, which is the more important thing for an initial project is a small number of people that really love something, not a lot of people that like it a little bit. And so although the retention numbers of the kind of internal ChatGPT prototype were very low, the people who liked it really used it a lot and really liked it. And I thought that was a very good sign.
And the first, you said two different things, right? The first was being embarrassed. Like it's, if you're not embarrassed about your first product, then you're too late, right? Are you okay with being embarrassed even today? You're okay with being misunderstood, and you're okay with being embarrassed.
Yeah. I mean, look, I want us to put out great products, but I'm happy for us to put out some early sketches to get feedback and learn, and not all of them will work.
But nothing changes with the scale of your company now because enough users trust you for...
I mean, things that we put out to like all of the entire ChatGPT user base have to have like a high threshold. And also now that companies, I mean, this is like getting close to critical infrastructure, companies are really reliant on this. And so for, you know, for enterprise products we ship, there's like a very, very high bar.
Okay. So now things have changed now.
But then we can still experiment with other, you know, I think hopefully you'll see us experiment with many new products, some of which will really delight you and make it into the core experience, and some of which people will look at and say, you know, it seems like cool in theory, but it doesn't really fit for me.
Interesting. And the other thing is we're talking about being misunderstood. Do you think you're still being misunderstood today?
I think at this point most of the world looks at AI and says, okay, I believe this thing is going to happen. Now clearly it's not all popular. And I think our vision for the world and what we hope for, how people will use this tool, this technology, is still not very well understood. But I think people at least understand AI is happening.
Okay. Interesting. Let's come back to you trying to do something extraordinary, right? And you taking bets on other people as well. And you're talking about YC, right? There's another quote which is, which you have only written, is I think you're some, I'm paraphrasing, and it's something like your, you have like your will to bend reality or your will to bend the world. The percentage of that particular thing is really high. Most people don't even try. Something like that you said. Do you remember?
I remember something like this, yeah.
Like most people don't even try.
I think it's more like bending the reality. You can like, you can do it much more than you think, or like people can do it much more than they think, and most people don't try enough. Something like that.
Yeah. And why do you think people don't take the moonshot?
I have like a bunch of theories on this, but many of them boil down to like people want to stay part of the herd, and it's like more comfortable to not do something unusual and maybe fail or maybe look wrong or maybe look dumb. And there's like a real, it's easy to point to the evolutionary biology reasons of why this would be true, like you don't want to get kicked out of the tribe or whatever. But I think people have like really strong pressures to not do that kind of stuff just because of the way society is built.
I mean, even more than that, I think society is built the way it is because of like evolutionary pressure, and I think that's probably in like very deep for most of us.
And what was in you? What was there in you which helped you bend this will or the world according to your will?
I think like a lot of tech people, I've always just had, which would be a negative in many other contexts, like a little bit less of whatever that internal pressure to stay on the track and do the same thing as other people. And I kind of think that's probably like mostly a genetic thing.
Just genetic? You don't think there's any other thing that you're building?
Oh, I'm sure there's other things too. But like, you know, you can find like all these studies about how people that are like slightly autistic are like tend to have, are more likely to sort of invent new science or like, you know, just get obsessed with one idea and work super hard in the lab against all the odds for it.
Yeah. And what was the will-bending moment for you?
I mean, probably starting OpenAI. Like, it's hopefully, it's too early to say, but it's, I think that'll probably...
be the most important thing I ever did.
What's your biggest worry?
Um, I mean, we are going to do something — or we have done something and will continue to do something — that is going to be an extraordinary force in society. I believe it can be a force for good, and it will be a force for good on net, a great good, but you know, it's like very high stakes and we have like a lot of responsibility.
True. You know, this takes me to the point where you wrote an essay which was in the beginning about SMI. Yeah. Right. And after that essay, during that May, you wrote to Elon about it and you said something that it would be good if someone else apart from Google started it, right? So you said like the biggest threat for continued human existence, the biggest threat is going to be SMI. Right. Why do you think that if you started it, it makes less of a threat?
Well, I believe in pluralism. I don't think only us having it is good either. Um, there are like — I think there are two kind of cartoonish ways this goes really wrong. Okay. Um, one is that there's like a loss of control to AI, and we intentionally or accidentally sort of seed control of the future to systems that are unbelievably more capable than we are, and we can't get it back. Um, and the other is like one person or company or country gets way too powerful relative to everybody else. Um, and I think that's very unstable. I think it would actually be much better for the world for there to be multiple AI efforts with slightly different approaches that are competing with each other and putting it out in different ways. And you know, we talked earlier about like the dislike of powerful people. I also have this like dislike of too intense power concentration. Um, I think that's not what works. So um, I'm hopeful we don't do that.
So is that what you were thinking when you started it as well? That someone with all the means in the world should not get this power today.
Well, at least there shouldn't be one. Like you want, I think, many efforts.
Okay. And let's say now there are like too many, right? There are four, five large companies. They have all the capital in the world. They have the great talent. They have all sorts of technology research. They all are working at it. What's your unfair advantage in this?
Um, I think we have, you know, the most users and the best stack from chips to data centers to models to other things that come around that, um, to sort of put abundant intelligence into people's hands. Uh, I think we have the best brand and the sort of values that people who are going to use this technology want. Like we want everybody else to succeed, and I think people can feel that from us. We don't want to be like the one company to rule everybody. This goes back to our founding.
No, I was talking about what makes you the Sam.
What do you think is your super advantage?
