저자: SamAltman · 외국인
요약 보기
영상은 AI 발전으로 스타트업 환경이 급변하고 있으며 혼돈 속에서 빠르게 움직이는 것이 기회라고 주장한다. 지수적 발전(모델·기업·사람 모두)을 신뢰하라고 강조하며, 에너지·로봇·칩이 AI 인프라의 핵심 병목이라고 언급한다. 또 기업 가치 창출은 주식회사 제도 같은 인센티브 정렬에서 비롯됐다고 설명하고, AI 권위주의 대비 분산된 권력의 중요성을 주장한다. 개인 경험과 장기 철학을 중심으로 한 인터뷰로, 특정 종목이나 자산의 가격 방향 예측은 포함되지 않는다.
다듬은 글
Today I'm sitting down to talk with Sam Altman, co-founder of OpenAI. A long time ago, you gave a talk at Stanford focused on how to start a startup. What has changed the most in the last 10 years since you gave those talks? Clearly, AI, what is now possible to do with a small team and the speed at which it can be done.
And not only what is possible, but what you have to do to be competitive given how much the world has changed. It feels totally different. I find it amazing how a 10-week-old startup can look today. And also, if you look like a 10-week startup from 10 years ago, you're in trouble. If you think of a 10-week-old startup today, what's the best example of one that moved extremely fast?
I don't even know their name, but I came across a startup that's about two weeks old, something like that, that has a completely revamped office productivity suite . Everything is designed for a world where AI needs to be able to use documents, presentations, spreadsheets, you name it. As a first-class consumer, and I don't know, I felt that would have been a year's work for a startup not too long ago.
I would assume that if the barriers to starting startups and the tools available become more common, it would be assumed that difficult startups become easier to start. What it means to be a difficult startup is changing so fast that I don't think I have a perfect mental model for what it takes to build a very successful company over many years.
What are the truly difficult and truly valuable things? I've heard a lot of people say, "Well, everything in the physical world is extra valuable now because software is going to become free." But, you know, it won't be long before robots are really good, and many of the expectations that it's very difficult to make rockets or whatever, that may change drastically.
I love times like these. I believe startups have their greatest advantage when the landscape changes the most, and when costs and cycle times come down rapidly. That's when startups, I believe, simply have a massive inherent advantage. And that's happening in so many places at once right now that it seems like a great time to create startups.
And yet, most startups say, "I'm going to create AI agents for vertical sector X." Well, that will actually work in many cases. It will be very competitive and perhaps it won't be, in fact, I would say it probably won't be the most successful startups that define that era. Hey, but it'll work. But, given how much the landscape is changing, I'm surprised that more people aren't doing that "I'm going to take on something crazy" thing with this completely new set of tools .
And also be willing to really internalize the fact that the laws of scale will continue and plan for things that aren't possible or economical this month, but will be in two years, four years, or whatever. So it seems like an amazing time to start startups and there's a very fertile ground, but also, and I'm not judging if this is what people are aiming for.
There is this incredible temptation to simply apply today's agents to easy wins, and I understand that. One thing you've said over the years is: you basically try, every time you meet someone new, to locate where they are in your mental model of who they are as people; And then, the next time you see them, you basically want to find out how far or how quickly they progressed.
Did doing that kind of thing, and doing it with perhaps thousands of people, influence your ability to say, okay, the model is this smart today, it's this smart 3 months later? Suddenly you can chart this and it makes it easy to believe in growth. I think there's something general behind both things, which is that I've simply developed a great deal of confidence in exponentials, whether in people, companies, or models.
Uh, I don't think, subjectively at least, that watching the model progress feels the same as watching a founder develop, but I do think the underlying belief system about the world is the same. If I were still advising startup founders, this is the most important thing I would try to make them understand.
And it 's obviously difficult for the same reason that there's still free money in the market to bet on high-growth young founders . And it's as if the market hasn't adapted enough for that yet. Well, I think the market hasn't adapted enough to the fact that the exponential progress of the models is going to continue.
And it's good to start working now on things that require smarter or more economical models . This idea of constantly changing chaos, how do you become good at it? How do you become good at operating in a chaotic environment? I think it's just practice. There are a handful of things that I believe, however much you understand them intellectually, require many repetitions in order to be able to handle them emotionally and operate amidst a lot of chaos, trusting that you will work it out and that everything will be okay, and that this is not what is going to kill you; You don't know how you 'll solve it yet, but you will.
I think this is something you can only learn by experiencing it. Um, and I actually think this is a real weakness of young founders: they haven't had the professional experience of achieving emotional peace with this. That's why it's so difficult for them at first, and although they eventually learn it, it's through a lot of pain, cost, and avoidable mistakes.
Definitely, if you end up in a high-impact, high- stress job, you'll learn to deal with the chaos and function very well despite it. At least most people will , but I don't think it's something that can be taught. I believe that one can only learn. You said something interesting a long time ago, which is basically: the first time you experience an event that can destroy your company, you feel like the world is ending.
Then you get over it , and by the tenth time it's not nearly as bad, and you think, "Well , I survived the other nine, this one probably won't be so bad." You also had another idea where you said at some point, I think a year or two ago, that you came to the conclusion that things would always go wrong, so you had to internalize the enjoyment of experiencing pain or of not-so-good things happening.
How did you make that transition? I was just reflecting when you said this, about how when I gave consultation hours at YC, I could always identify who the new founders were by the way they reacted when something went horribly wrong, unlike the more experienced founders; I often didn't know how new a company was, but I could tell if it was someone in their first group versus a founder who had been at it for years by their emotional state when talking about a problem.
Um, I think most people think that the opposite of a bad experience is a good one, and they prefer to have the good one, because it seems more fun or more enjoyable or whatever. But if you think that the opposite of a bad experience is having no experience, then at some point in the not-too- distant future you will be in the land of inexperience.
Then you know that you can also be grateful for the bad experience. Naval Ravikant used to say something I loved: if you had a fast-forward button on the remote control of your life, your life would be over. So the boring parts, the bad parts, are much better than having no experience. It's all part of what's interesting and adds a kind of depth and emotional range.
And I don't know, I find it quite easy to be grateful for the bad days. When you think about difficult problems and what motivates people to get excited about working on them, how do they decide internally within the company what to tackle and when, and when to focus solely on their core competencies versus broadening their scope?
I believe that a clear mission and a deep understanding of the problem do a pretty good job of pointing out what you should do. Not everything will go well for you. Sometimes you'll expand too much or not be ambitious enough. But we are very focused on this; it will massively empower people. It will be, you know, unstable but wonderful, and it is extremely important for us that power in the world becomes more decentralized and distributed.
In fact, one of the biggest risks of AI that worries me right now is AI authoritarianism. And, you know, a small group of people or companies thinking they need to control the world. That would be very, very bad. But given this mission, we feel we need to figure out how to make AI extremely abundant, cheap, and powerful, put it in everyone's hands, and have plenty to use.
And when analyzing the obstacles, there are a lot of new limitations such as chips, energy, data centers, robots and all the pieces that need to come together to build this platform and thus enable this mission. Now, there are many things you can build on that platform, like building each vertical, each startup.
We have no desire to do that. I believe a decentralized economy is important and good, and I would like us to be really excellent at producing these intelligence units and make the world think, "Wow, this is incredibly capable, an extraordinarily good offering," and we're going to integrate it into every product and service.
And that's why I think that's very instructive for us about what we need to do . Interestingly, the key inputs for extraordinarily abundant, high-quality, and low-cost intelligence— many of those things like energy and robots—are also things you really want in a world of abundant intelligence. Even if ideas are plentiful and all these great things are happening , we still live in a physical world and we still want things to happen.
So, you know, we need to be able to make things happen in that world. And I've wondered if the fact that energy and robots are going to be so important to keep up with this pace of intelligence infrastructure, and also that the things you need immediately afterward are energy and robots... says something profound or is it just the most obvious and boring thing, which is that to produce anything, including intelligence, you need to manipulate matter.
But it's interesting that we need these new areas so much for what we're doing now, and then much more immediately afterwards. What do you think is your biggest bottleneck? If you simply say that we want this to continue escalating without end. What is the biggest bottleneck you can see? Transistors and then electrons, in that order.
So, let's say, you take someone like Jensen, who is incredibly good at aligning all of his suppliers towards his vision of the future. I think you 're extremely good at this too. How do you plan to not only keep OpenAI on OpenAI's timelines , but also keep the rest of the world on OpenAI's timelines ?
You talk to them a lot. You have to do it, and not just say "I need you to deliver this turbine or this chip by this date". Because, you know, people will say, "Okay, whatever." And they have a lot of other priorities. You really have to show them: here 's the next model. Here's what it allows. Here is our research.
For our main suppliers, we really show them a lot about what we're doing, why we believe what we believe, and what this is going to enable. Well, you know, you have to decipher it ; I mean, the most important thing is to get them to believe in the mission and why they should prioritize it, but you also try to see how to align their incentives with yours as much as possible.
Charlie Munger has this great quote in which he says: "Every time I've thought I understood the power of incentives, I've underestimated them." How do you successfully align all the incentives of the different partners , as well as those of the people within OpenAI? Well, businesses are incredible vehicles to get started with this.
You know, as a child I was very fascinated by the industrial revolution and I understood it as a group of technologies that emerged at the same time and that, for some baffling reason, expanded almost simultaneously. And I would talk to my friends, my parents, or whoever about whether this technology was the most important invention, or this other one, or that one.
And my interpretation from my current perspective, as someone who has now spent more time thinking about business, is that the important invention was the limited company. And before it was invented, you had these family businesses that operated on the basis of trust, you had to know everyone, and the idea of a corporation didn't exist.
The idea of shareholders didn't really exist. And then, suddenly, this new thing was invented. Countries, like the sovereigns of the world, granted this type of new status, something that didn't really exist before. This powerful thing, we are going to let these new entities emerge and not give them the power of a state, but give them much more than the power of a person.
And fundamentally, this was about incentive alignment , liability protection, which is another type of incentive alignment. The ability to accumulate capital and suddenly you could do things beyond what a family business could do. And that environment allowed people to raise money to develop technologies.
Of a highly speculative nature. Pooling capital, finding ways to make different companies specialize in different ways and interact with each other. Serious financial systems and new types of instruments. This was truly something incredible. And the idea of a company has done a great job of aligning the incentives of very large groups of people.
Now you know, millions of people can be shareholders in a company. We do many other things, but I don't want to downplay what this incredible invention has been and the simply ridiculous performance of capitalism in human society. For example, one graph that I think people should look at much more than they do is the one showing the fall in extreme poverty over the last hundred years.
Like infant mortality and everything else . Any of them you want, any of the associated things. You know, if you look back, that's the most detailed view of the last hundred years. But if you look at economic growth, improvements in the quality of life, and all these things, and you were to plot all of human history by putting a line at the moment when business was invented, you would see a rather interesting change in the shape of that curve from then on.
Beyond that, I believe we have a mission that transcends any amount of economic power that anything else could achieve. And that's probably more important. One of the things I find most interesting is, have you ever seen those statistics that say what percentage of CEOs are sociopaths and psychopaths?
It's very, very tall. People are talking about this. So I wonder if capitalism is the first system that allowed these people, who are super selfish, to see their incentives realigned with those of society, creating great products and suddenly you help society but you also win . I know many sociopathic CEOs, no doubt, but I would say that the best ones I know wouldn't fall into that category.
There are many CEOs with very high egos, people with incredibly high opinions of themselves in a way that might have simply been annoying; But now, since people can, you know, buy his shares, maybe you tolerate his annoyances and eccentricities a little more. So , there's definitely something that aligns them there.
Don't know. I think there are many CEOs of very large companies whom I wouldn't put in the sociopath category at all. What do you think motivates those people? I mean , it's different for each of them. As definitely seeing how good they can get at the game or how much they can surpass themselves each year, I think that's a big part of it.
You know, it's something intellectually very stimulating. Mhm. For all other problems, it's like you're playing the most interesting strategy game, and that can be fun. If you had to think about how you work today versus how you would have worked 10 years ago if you were in the same position as now, what are the biggest differences between those two people?
Well, regarding the previous question, what motivates me most is that this is the most interesting and important thing I can imagine doing, and now we are at the singularity, this is the moment. For AI 10 years ago, this was at best a kind of distant dream. Hey. It seemed highly improbable, and now we are truly in the moment we used to talk about at the lunch table in a rather unserious way.
So what motivates me is that I've waited for this my whole life and I think it will be incredible, extremely positive and wonderful for the world. Uh, I'm excited to start working on it. I also think some of the alternative visions put forward by other companies are quite frightening . I will make sure that this is combated and that it is not what happens.
Uh, but the main thing that's different from 10 years ago is that we're actually in it now . This is real, I think it's true that everything is a crazy exponential and that no moment is the turning point, but also that somehow we are in another of those decisive periods where the curve can go one way or the other, like when we started 10 years ago.
What do you think are the biggest impacts of doing this part of the curve well? I think there are still significant AI alignment issues to resolve, and security issues, and there are still important ones. Economic issues, the future of work, and all that kind of stuff. But I believe that the struggle of the present moment is.
Are we heading towards a world of AI authoritarianism or freedom? Are we going to decide that, because of the very real security and economic problems, we want a single model to be the machine god or the company associated with it? Or are we going to say, "You know what? It might be a little chaotic, but every time humanity has sacrificed its freedom for security, it's been a net loss in the long run," so are we going to put this in the hands of the people?
We're going to give them power. We're going to let society express its ideas and use this technology in whatever way it wants. Of course, under certain safety regulations and some limitations, but you know, with a lot of power and potential. Oh, and it really matters to me that we do it that way. When you were talking to Patrick Collison on stage a couple of months ago, you said something fascinating: that you basically talk to or message 300 or 400 people a day internally.
I send messages to more people than anyone I know. I don't think this is a good thing. In fact, I think it's a bad habit. Because? Because it takes up most of my time and now I have a very short attention span. Interesting. What allows you to do that which you otherwise couldn't? A huge amount of context.
I mean , I'm like an AI model that isn't that smart, but has a lot of short-term context about whatever is going on at the moment and then I forget it. Does it allow you to make decisions that others wouldn't make, making the company's decision better or not? Don't know. I mean, it allows me to make different decisions than others would.
What different decisions have you made in the last 12 months that you wouldn't have made if you didn't have so much context? I do n't know if I have any specific ones that I can share publicly, but there are many things where, because I know about some upcoming research breakthrough that will affect a customer or supplier or whatever, I can make a slightly different decision at the time.
And if I hadn't known that morning, I would have made a different decision. It would have been worse. One of the things I think you're best at is long- term thinking, and yet, it also seems like things are moving so fast and the world is changing so much that planning 20 years ahead and then looking back, I don't know if it makes sense.
What are your thoughts on this? Is your time horizon getting shorter? Actually, I don't try to plan backwards. I try to have a small number of firm convictions about the future that serve as a direction for where to go. But then, I try to plan ahead from the current state, about what we can do now, what we can do this year.
And sometimes you do have to plan some things with a 5 or 10 year horizon, you know. But I try to plan ahead guided by a small number of beliefs about the future. I think there are many people who have too many beliefs about the future and a somewhat rigid worldview that they try to conform to. And then, what happens is that they end up, for example, chasing things, you see how space companies become AI companies or whatever.
Um, have only a small number of things you deeply believe about the future and be very flexible with the rest. And staying very true to the essentials is useful. One of the people I know best, his main corporate value is simply the critical path and focusing only on the critical path and staying concentrated on the key factors of the biggest obstacle.
Simply solve that, then move on to the next one and do it again. How do you think about the critical path in your own life when you try to make these kinds of decisions? I mean , for so long I've felt focused on this singular goal of abundant intelligence and the belief that incredible human prosperity will come from that, as long as we do n't have a strange concentration of power and a new kind of authoritarianism.
For me, it has been quite clear at any given time what is on the critical path to get there. And I haven't been tempted to, you know, reconsider the goal or think about these other things. I've been thinking about that a bit more recently, like, well, if we're really close to superintelligence, what's next?
But that's how I subjectively feel we're close; it's the first time in over a decade that I've thought about what's next. And is it the ranch? Uh, I mean, eventually there might be some other things down the road, but that's, yeah, that's the eventual plan. If you had to try to guess what else Sam wants to accomplish before he goes to the ranch.
Uh, look, super artificial intelligence is going to be built, you know, it 's not finished, but we're on the gliding trajectory. The path to get there is abundance and decentralization, and broad access to it, ensuring that it results in widely shared prosperity. I would really like to achieve that.
So maybe that's what I've been thinking about next, how we're going to accomplish our mission technologically. I think that much of our worldview and our beliefs about security and the impact it will have are things that we have also achieved. But you know, widespread shared prosperity and a real focus on empowering the world here.
That seems very important. I've heard you say this idea of getting on planes in extreme situations time and time again. I've done it for you so many times. So far it has worked quite well . Good. I think it's very solid advice. When was the last time you boarded a plane in a life-or-death situation?
Well, I can't say what for, but it was very recently and it worked. Congratulations. Thank you. And I really didn't want to. Those were two very inconvenient nights away. I didn't want to do it. New baby, a real mess , but I did it. Alright . Take me through your thought process for deciding to do this world tour.
I mean, it 's hard to go back 3 years because now it seems like AI has always been here , but 3 years ago we had just launched GPT-4. The world was falling apart. Everyone was panicking. Everyone was like, "wow." And nobody knew what to think. And I could feel these storm clouds gathering and people getting angry, or it was like world leaders were saying, "Do we need to take control and shut this down?" It was as if this was waking up; it was a strange moment, and because it all happened so fast, people didn't have time to process it, go through this process, or whatever.
So, you know, we were getting all these things escalating, and all these world leaders were saying, "Will you come and meet with us?" "We're thinking about doing this and that." And I was like, "I think if I, or someone from the company, but at that time the company wasn't very big, so I thought it would probably have to be me who showed up and talked to people." I had a feeling things were going to go very badly and I was thinking of taking a lot of short trips, but I hate long flights.
I hate jet lag . I had all that. So I thought, "I'm going to finish this in a short period of time." Brian Chesky advised me to do it. He had done something for Airbnb, something similar, but he had visited like eight cities or something like that. And you said, "To hell with it, let's go big." "Let's do 30." 20.
I think we visited 28 countries, 35 days. Uh, I practically lived on a plane. It was very strange. Did you have a bed? I had one...yes, I was comfortable, but yes, I had a bed. Um, I fly economy class, so it's a little different. This is better than that, right? This is much better than that . But, you know, traveling is still a hassle.
You can make it as comfortable as you want and you'll still miss your own bed, your time zone, your office, and everything else. One thing I've learned is that no matter how many people have those cures for jet lag, they're all lies. It's simply terrible. Your brain doesn't like it. Never, because I went to so many places.
I think my biggest time zone change was 4 hours for the entire trip, and it was mostly one- hour jumps at a time. So I didn't have as much jet lag. I was simply exhausted and it was very strange. It was great to be able to see the world so quickly, because you really pick up on cultural differences that are very subtle when you go from one place to another.
But in the end, in the last few days , I started having these weird dreams, you know, between asleep and awake. But I started to have this vision, feeling, or whatever, of being in my childhood bed. Interesting. Something I had never had before or since. But I would wake up vaguely and think I was in my childhood bed, in my boy's room.
And I thought, "Okay, this is something deep, it's time to go home." It's like your subconscious is fighting back or something. Something of the sort. Interesting. When you're in zombie mode and trying to function, or in goblin mode, whichever you prefer, how do you change ? What changes within you?
How do you decide to move forward? I definitely reached a point near the end where I was counting down the days until I could be home. Well, it's just that every day was, you know, like 14 hours nonstop and I'm not an extroverted person, you know, I would meet hundreds of people a day in some cases.
Well, tensions eased in the end, but at first the world was very nervous and then, you know, I think once it just calmed down. I think something that helped was that I was clear about when I was going back home and I knew I would have to do it again for a while. Well, I've taken shorter trips since then.
I'll probably do something like that once or twice a year. But I was going to say that something I learned is that I much prefer grouping several international trips into a block of 7 or 10 days rather than taking a bunch of scattered day trips. So I always try to do it that way. Now I think many people overestimate the risk involved in most stocks.
I mean, there's a lot of risk if you're just, you know, buying a bunch of stock options on Robinhood, but there's probably a lot less risk if you're getting on planes and stuff like that. What are some of the best examples of times when you basically did something that seemed very risky to others, but that you knew or believed wasn't?
I suppose the obvious example would be that I always had an extremely high conviction when buying many computers. Have you ever seen the meme that says "everything is high risk if you're a coward"? No. Good meme. Well, I spend a lot of time trying to talk to people about why they think a specific decision is high or low risk.
And I 've discovered that simply getting people to talk about it out loud often at least overcomes their intellectual blocks in either direction, because people usually get it wrong one way or another. Not always in the emotional aspect, but sometimes that too. If you had to think about one of these meetings where someone is strongly against a decision you want to make, what is the most effective way to help them see what you see?
To be honest, I don't think I'm very good at this. I think the right answer is to spend a lot of time really trying to explain it and make people understand it. And in fact, I think I used to be better at it. And since life has become so hectic , I just get more annoyed or frustrated , or I say, "This is what we're going to do." And I don't think that's a positive characteristic.
And I'd like to return to that idea of: let's analyze it properly. You said you used to be much more difficult to work with, I don't know, 2008, 2009, 2010, around then, than you are today. I. I do n't know if people would describe me as someone fun to work with. I like to be someone interesting to work with, effective, who increases ambition, very mission-focused, but in the day-to-day, really fun, I don't know.
I don't think I'm a particularly funny person in general. I feel that if you're having a good time, that tends to be contagious. And I think if you generally work on things you want to do, that's having a good time. Perhaps I'm wrong. I mean, you could ask a group of people I work with. I don't think so, I do n't think you'll get a resounding "he's fun to work with." Perhaps not in those words.
I think people would say things like: creative solutions to problems, new perspectives, a very high level of ambition, it made me do something I didn't think I could do, that kind of thing. What is it usually like when you walk into a meeting with someone and realize they are not ambitious enough and you need to try to help them become more ambitious?
The version of this that came to mind when you said that was during the first sessions with new founders of YC; You take an average person who has worked at a corporate firm, you know, a big tech company for a few years, and the level of ambition, scale thinking, self-confidence, whatever , is catastrophically low, horrible, horrible , and you realize that person has never been without a boss in their life.
They have never lacked a father, a teacher, a manager, or someone to tell them what to do or what they were allowed to do or whatever . And they were also punished for being too creative, wanting too much , thinking too big, or being too ambitious, whatever. the. Most countries or cultures have some kind of phrase for when someone is too ambitious.
Basically. The tall poppy syndrome. That's one. Well, and you realize how deeply ingrained this is in people, that they heard it from their parents, their teachers, their friends, or whatever, from a very young age. And getting people to simply stop thinking like that takes time. What is the most effective way for someone to go from not believing in themselves to believing in themselves ?
I believe it's through small, consistent victories. Like when you do something you did n't think you could do, or you try something that seemed, you know, too ambitious or with very high expectations, and it works, then you say, "Okay, try it again." Do you do something like Steve Jobs, who if someone was worried, would look them straight in the eye and say...No, nothing like that.
I do n't believe it. It's more like a "you can do it". I am not, I don't, I don't give inspirational speeches. No, I don't put any pressure on anyone like that. What is your favorite strategy game? It's been so long since I've played anything other than my poker night with friends. I wish he would. I guess I'll have to say poker, but I...
I miss good board games. What was your favorite board game? There isn't a single board game that I say, "I want to play that over and over again ." I'm not someone who says that Catan is the best board game of all time or anything like that. But they're all fun because they're new and you have to solve a new set of challenges, rules, and everything else.
When you think about the business world, I believe you build your company differently than almost anyone else can or is willing to build companies. What does that allow you to do? I actually think that every large company is quite different. I don't think this is a unique situation. I think if you make a list of the 20 largest technology companies or whatever , they are all surprisingly different from each other in terms of strategy, how they operate, and culture.
I think we're quite different from everyone else, but that's no profound revelation. What would you say is your superpower? I think we did a very good job maintaining a firm conviction in something that no one else believed in, even though it was the obvious thing one should believe in, and then putting together all the pieces and the talent necessary to make it happen.
What allowed you to have that conviction when other people simply didn't have it? Mmm, I've thought about this a lot. Do n't know. It seemed so clearly obvious to all of us that I was more worried that we were believing our own lies than that everyone else was wrong. But with the benefit of hindsight, we were clearly right.
And I don't understand the deep mental block and lack of conviction of everyone else. It's very strange. I've spent a lot of time trying to understand this. Did you find any explanation in your personal search for why or not? Hmm. I mean, there are all the obvious things about groupthink and that big transformations don't happen that often and exponential curves don't stay powerful for that long that often .
But, you know, maybe between 2016 and 2018 we had to be very lucky and it required a lot of faith , but by 2019, and certainly by 2020, we really shouldn't have existed. Google should have kept everything. Yes. I think Jeff Bezos called it a business miracle: that they built AWS and then had no serious competition for seven years.
And I feel that OpenAI is another business miracle. Clearly there's something here about why large companies become sclerotic and stagnate, and that also helps them in many ways. But, I mean, this is wonderful . I think it's great that big companies won't remain the only dominant forces forever. That would be really bad.
But clearly there is something where these business miracles happen when they shouldn't. Is that okay, or should it happen? Well, I think it's great for the world that they happen, don't you? But to go back and explain in 2019 why the giants, with huge amounts of capital, talent and everything else, allowed OpenAI to do what we have done.
It's difficult to explain. Why did Microsoft give them the money? Hmm. I think that's easier to explain: Google had DeepMind and Microsoft didn't have a stake in AI, and you know, big companies feel they should have something at stake in all the major areas of technology. Was it something like they thought it would be big, but underestimated just how big it would be?
I think it's been great for them. I don't know if we've added $1 billion, $2 billion, or whatever to Microsoft's market capitalization, but I think it's been a lot. And I think they've achieved incredible technology and growth in the cloud, and they will continue to do so . Yes, I'm sure it's bigger than they thought it would be, but I think they made a good bet and it's worked out great for them , I'm very happy about it.
There have been different versions of OpenAI. The first was the research laboratory. The next one was the product company, and you said that, like, you added the product company to the research lab, which is exactly the opposite of what most people do. And now you're basically going to be, like, a massive infrastructure company, and, like, almost all of OpenAI's value will come just from this low-margin infrastructure.
And you also said that basically, to transition to the next phase of this business, it's not necessarily something your brain is naturally designed to operate on. What do you need to change, and what is the next phase of the business? I'm not responding to this because I don't like it. I think this is getting into something direct now, something we're not ready to talk about.
But I'm very excited about the next phase, and I think I've figured out how to align something I 'm very good at and passionate about with what will be our next 10x or 100x growth. So, what was the transition like for Sam, the head of the research lab, to Sam, the head of the product company? What were the biggest transitions for you in that?
I mean, everything was extremely different. It has felt like two almost completely disconnected jobs. And it's not like I did one and now I'm doing the other, because I'm still responsible for the research lab. But they are almost separate things, and definitely running the research lab did not prepare me to run the product company in any way.
Did the YC experience prepare you more? I had seen many people have to climb the ladder in large companies. Was it fun to finally get behind the wheel? No. Um, I... The fifth day after the launch of ChatGPT was the day we surpassed one million users, and I had seen it go up on the first day and then down in the afternoon, then up on the second day to a higher peak and then down, up on the third day to a higher peak and then down.
And each of those days the researchers would say, "Oh, that was just a passing PR craze ." This is over and you know that's not going to happen. And I had seen enough at YC to know that when something grows like that in a completely organic way, it had, for me, all the spectral signature of something that was going to happen.
And then, on the fifth day, it surpassed one million users and I got home, and that's when I really realized that we were about to become a company, that it was going to go very fast and that it was going to be crazy, like we were going to transform into a big company very soon. I had been watching what these founders were going through, and when I got home, Ollie said, "Oh, I saw that you've surpassed one million users.
Congratulations." I told him, "You have no idea how bad this is. You have no idea what's about to happen. It's not just bad for me, it's bad for you too . We have this quiet, nice life, you know, it's really wonderful. We're about to get shot out of a cannon." And that's what happened. But it was a very abrupt transition.
I knew it was coming, and I'd seen other people go through it, and I knew it wasn't a pleasant experience, but I thought, well, we're being shot out of a cannon. Here we go. Were you subconsciously preparing for years, knowing that this would eventually happen and that your intelligence would be good enough to produce a magic product?
My opinion is that I intellectually knew it was going to happen at some point. I had no idea when. And I did n't know that would be the moment, of course, but I had decided to pretend to myself and had successfully fooled myself into acting as if it was n't going to happen. Running the research lab was, honestly, the coolest, most fun, most amazing job and lifestyle I could have ever imagined.
In what way? I want to I mean, it was like it wasn't stressful at all. It was intellectually incredibly fulfilling. It was the smartest group of people doing probably the most important work that's happened in the last century, maybe longer, I don't know. And I had a front-row seat. And that was a once-in-many-generation moment .
It was amazing. It was the coolest thing. Are you guys experiencing that same kind of growth? You know, I keep seeing Tibo posting on Twitter that basically here's another reboot, here's another reboot. Are you seeing that same ChatGPT-type experience again? Absolutely. I think so far there have been two big form factors and their associated growth.
There was the chatbot type, and then the programmatic agents, and the programmatic agents are going crazy. I think there's going to be a third one soon, which is this idea of persistent agents, chiefs of staff , coworkers, colleagues, whatever we call them. And you know, I think that's coming pretty soon.
So we're going to go through the third of these waves pretty fast. A long time ago, apparently in August of last year, I saw you post something And I thought, this isn't going to help the reputation. Um, let's see. Why? Why did you do that? Um, well, first of all, I love that scene in the movie. Like when the Death Star comes out of hyperspace and that dramatic music is playing , it's a great scene.
I thought it was funny, I was, you know, scrolling through Twitter. I think it was late at night. I don't really remember. This is where conspiracy theories come in . Like, you're adding fuel to the fire. I just thought I was having fun at the time. It was, I do n't know. To be honest, it wasn't a great tweet, just so you know.
Actually, I thought it was cool. I thought it was funny . I thought it was funny. I think I thought it was funny. Yeah. I feel like the world is feeling that way more and more . Uh yeah. Uh yeah. Okay. Yeah. It's a good meme. The way someone explained the marketing challenge for OpenAI and the industry in general is, number one, we're close to creating what we need to explain to the world.
Number one, we're close to creating a A genie who can grant any wish. Number two, we're going to make sure that our first wishes broadly benefit humanity and take the world to a place where many more people can have many more wishes. And number three, the space of what you can wish for is incredibly vast and creative, and it will be exciting to explore it with real human values and preferences .
But, but I think there's a fourth thing as well, which is exactly that. You start making these wishes, the computer grants them, and then you think, I didn't think that was going to work. Yeah. Now what? It's a strange feeling. Yeah. It's like the problem where I guess, you know, people spent like a hundred years trying to disprove some, what was it?
Jacobian. Yeah. This thing. And then someone just asked Claude and it was like, I don't know, in a few days or something. So , I think there will be a lot of great jobs in the future. I really do . I think we'll have a lot of intellectual satisfaction. And I think most jobs are going to adapt more than it seems like they should.
But math, I think, is a very important example. So we can study something closely right now that might not go that way. In fact, I'd say it probably won't. It seems that even as the systems keep getting better , the people I know, nobody is less busy. Everyone is busier. Everyone is doing more things, except for those who say, "Oh, I spent a billion tokens ." "Hey, so what, I love this thing I saw where someone said, 'I spent, you know, an unholy amount of tokens.'" "And then someone said, 'Why aren't you more successful?'" "You know, technology has long promised people that they'll work less and have all this free time.
And it's been going in that direction. I think people do have more leisure time than they did at many previous points in history and a better quality of life in many ways. But somehow we never delivered on the promise of the 4- hour workweek on a massive scale in society. And I do n't expect AI to change that.
Uh, clearly society isn't working for a lot of people, and more productivity gains that actually benefit people would be a great thing, and I suspect they will happen . I suspect that will happen. But something I really admire about people is that our expectations keep rising. We always want more. We think about new things to do, things to create for others, things to desire for ourselves, and, you know, it's kind of a relative game.
People are very focused on how they're doing relative to others. And so, the competition that drives the economy and the kind of, I think, wonderful desires to want to be useful to other people, to create something and serve, all of that as a whole. I hope that continues, and I think we'll all be much busier than we thought in a post-superintelligence world, and we'll still complain about it, but secretly we'll be happy.
How do you think the status games change as less of our direct input correlates with value creation? I don't think it will feel like ours at all, like the things we value. I suspect it will be things that are very human. Well, one thing is, for example, cooking for others. You're not creating great value in the world, but it shows love.
You're creating value in the sense that they 're experiences; you're not creating economic value in the traditional sense, perhaps. Um, cooking for others is a great example. I suspect that people cooking and eating together will still be important long after robots can cook very well. Do you think a good heuristic would be almost like the greatest works, the ones that last the longest?
Take the Bible, for example; it will most likely still be impactful a thousand years from now because it has been impactful. over the last 2,000 years. Do you think if you had to extrapolate what things humans will continue to care about, it's like, what's the most primal and earliest thing? You know, we love adventure, we love a good hunt , cooking.
Betting against evolutionary biology is generally a bad bet. Yeah. So I guess those things will continue. When was the last time you realized you weren't being ambitious enough? I mean , I definitely grossly underestimated investments in computing. Could you have known that beforehand with the right mental model ?
Yeah, but I got swayed by the financial markets or something. I don't know. That was clearly a mistake. How do you correct that going forward? Well, I hope to learn from it and I'll make new mistakes, but I won't make that one again. As far as computing goes, it's like the biggest infrastructure project of all time, or it's about to be.
Yeah. Are you going to try to vertically integrate everything under the OpenAI wing , from From power generation to token generation and, you know, serving a person on ChatGPT or Codex or something like that, or are you going to try to have a bunch of partners? We're not going to try to do everything literally within OpenAI, obviously, but we will try to do a better job with a well-functioning supply chain than we have so far .
As you know, we've brought together chip design and model design, and that's been good. I don't think we need to bring together electron production in the same way because that's more of a commodity. However, when you really think about the whole supply chain that has to come together to produce intelligence, I would say there's relatively too much focus on algorithms that create better algorithms and not enough focus on data centers that can create more data centers, which, in a world of robots and a truly automated supply chain , is totally conceivable.
You can dedicate the processing capacity of a data center to running a fleet of robots to make more copies of the data center. And that 's probably a wonderful thing to do. In relation to other uses of that capacity Computing, if we're right about what that additional capacity will ultimately unlock.
How much time do you spend thinking about how to successfully design that fully robotic, end-to- end future where intelligence can simply produce more intelligence? The time spent thinking almost all goes into execution ; the idea is a very obvious idea, it's very easy, you can say relatively quickly: here are the pieces that need to be put together, but then getting all that machinery and all those companies to work together, that, you know, doesn't have the glory of the big idea and trying to think of big thoughts, but rather a lot of hard work .
When you think about execution, what does that normally mean to you? No, there's nothing typical there like, I mean, to the extent there is , it's simply taking whatever steps are required in the problem and finding a way to make it happen. Mhm. But, you know, figuring out how to finance the construction of a new plant is very different from figuring out how to assemble a great chip design team and then get them working with the research team and then actually get the supply chain up and running.
Well, each of those is a very different approach. One thing I know Brian Chesky does is talk about shamelessly trying to find out who the expert is on anything and then basically going and asking them any questions you have, like you're trying to catch up on something. What's your process for doing something like that if it's something new that you need to get good at in a very short period of time?
Definitely find the experts and talk to them, reading as much as I can. Uh, I don't have a lot of big strategies, I don't have a lot of novel strategies here. I've always been very grateful and pleasantly surprised by how much experts are willing to help people if they ask. Uh, it's like a very nice thing about humanity.
One thing I've heard you say is that you should always ask for what you want. Um, because you don't always get it, but sometimes you do. Sometimes you do. And when you do, amazing things can happen . When was the last time you asked for something that would seem absolutely insane to someone else and you got it?
Actually, it was recently when I got on a plane which I can't talk about yet, but okay, back up one more. Back up one more that you can talk about. I mean, in a way , Codex was an example of this—this isn't the most recent, but it was one that I think is instructive. Um, we were way behind Cloud Code, and it seemed like a crazy suicide mission to try and beat them with a coding app.
And you know, the consensus is that these kinds of things never work, and you just move on to the next one. But you know, it's a waste of time trying to win when somebody else already has momentum in a particular product category . But we decided that we thought it was really important. We asked a team to do it.
They achieved something legitimately like incredible, very rare in the history of business, and now it's the product that most of the best programmers I know use. Um, and that felt like an impossible thing, but if we hadn't asked the team, "Hey, we have a really important but extremely difficult mission for you," it just wouldn't have happened.
Why did you make that decision? Why Did n't you just give up? It felt like one of those few very strategic areas that was going to happen very quickly, and programming is so important to RSI, not to mention the economic value, that we just didn't feel it was one we could let go of. YC, you basically had the most distilled version of what the mission was: just trying to increase the amount of innovation in the world.
My feeling is that if you really thought, "Uh, why did you decide to go work at OpenAI instead of YC?" "You could almost argue that you could have a greater impact on increasing the amount of innovation in the world by working in AI. It wasn't that intellectualized at the time. It was just that I knew—I mean, my whole life I wanted to work in AI, and I knew it would be the most important thing I would ever touch.
Like, for me, it was my passion. I just thought it would be really important, but I just wanted to do it . What were those first few weeks like? I know there was this—I don't know if you would describe it as a camp— but people basically got together for a kind of retreat, and it wasn't even meant to be a company at that point.
I think it was just like the first few weeks once we started. No, we were at Greg Brockman's apartment. Yeah. Um, and there were maybe 10 or 12 of us. And, you know, it had been all this work getting it going, and then we showed up one day. Mmm. The first day, I was like, you know, January 4th, something like that.
And it was like, well, here we are. What are we going to do? We should get a whiteboard. We should start..." to talk about ideas. Maybe we should write articles. And I had this, oh [ __ ], what have we done? It was like a really crazy moment. Um, you'd already committed to the world and a bunch of other people were doing it?
It took a couple of years to find our rhythm and figure out what we were going to do. It just wasn't, I mean, we knew we wanted to figure out how to build AI , but beyond that, man, really unclear. When you're in the stumbling-through-the-woods phase of this kind of, you know, Manhattan Project, how do you most efficiently stumble through the trees to quickly figure out what not to work on?
I mean, if something isn't working and you run out of ideas, you can kill it. That's an easier thing. The really hard thing is when something is working really well, when do you decide to eliminate the other things so that it works even better? Well, there have been a lot of pretty significant moments in OpenAI's history where something started working really well and we eliminated other good things so that the best thing would work even better.
So when GPT3 started working, we shut down projects like the ones in robotics, which we were very excited about, to really focus on this . And then, when the scheduling agents recently started working , we shut down things like Sora , which we were also excited about, and the browser to focus on this.
And that's hard to do. Um, let's say you invest a billion dollars in a new business like Sora. How do you decide whether or not that business will become , you know, a pillar of OpenAI? Oh, no. It would have been super successful. It was just more important to dedicate the computing power and energy to the scheduling agents.
Take me to one of those meetings, imagine you're in that meeting. I suppose maybe you were in that meeting where you said, "We've spent over a year on this, huge amounts of money, huge amounts of computing power, lots of people and resources, a lot of momentum, people are using it, they love it," and then you decide, " We're going to cancel the project." How do you make that decision?
Yeah. It's not a one-meeting thing. I think it's a somewhat gradual realization that there's a more important use for this computing power, these people, Um, you know, this product direction, and we're going to make a very painful decision to achieve it. How do you realign people once you have to, in a way, kill their baby to pass them on to another baby?
People understand the mission, what's at stake, and the need to reorient themselves to achieve it. So even if they're not happy in the moment—sometimes they're very happy, sometimes they say, "Yes, this is the right thing for the mission." But even if they 're not happy in the moment, they say, " I understand why we're doing this." And that kind of continually refocusing toward superintelligence is a good thing.
Do you think there will be current projects that will be eliminated? I guess it's harder to pursue. Yes. There's an amazing quote from John Collison where he basically said that everything in the world, if you look around, is so difficult to achieve. Even building a park bench is incredibly difficult.
So when you look around, you can think of the world as a universe of exciting projects. And I think if you're in someone's position like Steve's... Jobs, where you're going to create something that a billion people could interact with every day for hours. You have to think deeply about how to design something that people will love and, preferably, that won't make them unhappy.
How do you think about design? I feel very fortunate to be able to work with Jony Ive designing beautiful things. Hmm. I've learned a lot from him about how he really studies a problem before trying to come up with a solution, and he doesn't even allow himself to think too much about the solution until he really understands the problem.
And I think this is a key that I didn't appreciate before. Really great design is much more about understanding the problem than about a flash of inspiration. And if you try to rush into the object or you get too attached to it, you won't do as good a job. I think the iPhone is the greatest piece of technology that humanity has collectively created so far.
It's an amazing thing, but I don't like my relationship with it anymore. I turned off my notifications, so now I like it a lot more. Yes, I keep it on "do not disturb" mode . That wasn't even enough. I just turned it off Notifications for everything, except for a very small number of things. And I never have them turned on because, you know, I can look at it if I want to.
Even messaging apps . Even messaging apps . That was a huge improvement for my life. And I deleted TikTok because it was just too powerful. Uh, were you addicted to TikTok? You know, uh...something crazy happened here. When we were building the Sora app, I forced myself to become addicted to TikTok because I just wanted to learn.
I thought, you know, I'd never really used it before. I mean, people would send me a TikTok or something, but I never got really into it . And I thought I did n't really want to build something that would have that kind of effect, that would suck up people's time, but no, and I loved it. I love TikTok.
I really thought it was cool. Um, and then I thought I could control it. I thought, "Oh, you know, it's actually fun, but I only use it for about 10 minutes to relax before bed, and I'm totally in control." And then it was like an hour one night, and after One Saturday afternoon I was on the couch for like three hours, and I thought, this really wasn't what the iPhone was supposed to be for.
Now, I'm really enjoying it in the moment, like a drug, but I can tell it's bad for me. And then I briefly regained control and cut down to about five or ten minutes a night, you know, just something to relax before bed. Then I just said, enough. I think the iPhone is amazing, and yet I didn't feel like I had enough self-control to keep that app on, or that the notifications from messaging apps were a positive thing for me.
And I'm sure we're going to create an amazing, beautiful, really useful, and empowering set of devices. And I'm also sure that people are going to misuse them and that they're going to make people's lives worse in ways we can't even imagine. And we'll adapt. But that's the thing about powerful technology.
When you invent a new device, what's that process of exploring through the problem space? Like, Johnny, I don't speak for him. Um, but he's talked about this. Before, sort of, but it's like his process. You know, when he's making a car, he studies the entire history of motoring, and then the different typefaces people use for text in a car's cockpit, and the materials and why, and the different sounds engines have made over time and why some have been popular.
I mean, he writes literal books of all his explorations of all these incredibly small and detailed components. Did you guys basically take those books and try to use them as training data to create something like a model of Johnny? Um, no. But that would be a great thing to try. Yeah, there's definitely a part of his process I don't understand.
I understand the study, I understand the refinement of a very new concept into something great. But there's an intermediate step, the inspiration that comes from understanding the problem really well to arriving at this novel idea, which seems to just happen all at once, and I don't understand that.
When you think about designing something, what goes through your mind? I'm not a designer by any means. I I, no, I, uh, try to be pretty aware of what I'm not good at and I do n't try to have a strong opinion about it. What are you worst at in terms of building businesses? Maybe product. Really? Maybe.
Okay. How do you find great product people if you're bad at product? That's something I've never quite agreed with people on. There's this business meme that you can only hire people for areas you understand deeply. I think I can tell if they're great. I think this is clearly not true. I also don't understand design.
I know Johnny is great at design. There are a lot of things I can point out when you talk to him for 30 minutes. It's very obvious. What things do you actively try to improve on and which ones do you just try to delegate, understanding that you'll never be one of the best? I strongly believe that you should try to improve on your strengths, and that whole obsession with "I'm going to get better at the things I'm not good at and never will be" is a big trap.
Being super good at your Strengths. Okay. What are your greatest strengths if you had to rank them? I hate this question. Uh, I don't hate it out of false modesty. I think it's very difficult to say anything revealing about your own strengths. You're clearly good at uniting people around a goal. Sure.
But I don't think I could teach someone who isn't good at it to be good at it . I think things that come naturally to someone are very difficult to, I don't know. Or at least in my case, I don't deeply understand why I'm good at it or what to do about it . Although I think we talked earlier about this category of things you can learn but can't teach .
Hmm. I think this is quite important to understand, and if you want to become good at something like that, asking someone to explain it to you, in my experience, never works. Really studying them, sitting with them in meetings and trying to observe and learn it yourself, that does work. But when there's been something I've really wanted to improve on, I've tried to be around someone who's great at it and study it deeply.
But I don't think that they could have explained it to me. I don't think they could have taught me just by talking about it. I think that's absolutely right. If you look, for example, at how the best in the world learn video games: if you play, say, CS:GO, you play a bit, learn the basic mechanics, and then go and watch a pro, observe how they interact on the map at different points in the game, and things like that.
When you think about the pace of an organization, you know, there's ambition and all that, but ideally, over time, you don't want your organization to slow down; ideally, you want it to move even faster. How do you integrate that? I think it's 90% the people you put in leadership positions. Okay. There are other things you can do.
People have all these different operating paces regarding how they try to run the company, this management technique, or that one. And I think, to a large extent, it all comes down to the people. Definitely, something I think about with anyone in a position like that is: are they fast-moving or Slow?
How do you measure that without having actually worked with them in the past? Well, ideally, you've worked with them before. I think that, most of the time, a company's executives should be promoted internally, not hired externally. Agreed. But when you need to hire someone externally, you spend a lot of time talking to them, checking references, and trying to work together in some informal way.
Changing the subject a bit. What was the most painful thing that happened in the last 12 months? Honestly, having kids and working really hard at the same time is brutal. It's just extremely painful. You know, I think I'm a pretty present dad by comparison. I do nothing but work and be with my family at this point, but I still feel like I'm missing out on so much of this unique stage.
It's very painful. When you think of someone like Masa, for example, what is it about his brain and his way of seeing the world that's so different? I think he's going to end up being one of the biggest successes of this decade. Um, an amazing person. My dear friend, an amazing person, with great Conviction and faith, who isn't afraid of big numbers and scale at all, and that's great.
Uh, there aren't many people like Masa in the world, and we should be very grateful for them. What made him like that? Well, I don't know if you've spoken to him, but... Uh, I mean, as far as I can see , he's always been like that. Do you also have that same kind of concept of limitless scale? No, I think Masa is a one-of-a-kind character.
Do you think of OpenAI a bit like the goose that lays the golden eggs? In what sense? Um, generating a lot of very valuable things over time and building a core competency of greater intelligence in the world. It's not really how I think of it. I understand what you're saying, but it doesn't feel like we're just spitting out golden eggs.
It feels like we're building something that keeps accumulating. I think one of my favorite mental models that I've taken from you is this idea of real trends versus fake trends. And oh yeah, this is a big thing. Whenever people He was talking about bubbles or whatever, and I was a little confused .
Like, you know, I studied Warren Buffett and Charlie Munger a lot, and they love to talk about bubbles and overvalued things. I was trying to figure out how you discover if something is real or fake and if it's going to persist and continue or just revert to the average. And what did you discover? Well, I think it was your mental model of basically, if there's a small group of people, what's a real trend versus a fake one?
A fake trend is, you know, the virtual reality situation, where there's a lot of hype, but then someone buys the device, they don't really love it. They don't really start designing their life around the experience of using it, and then it just sits on the shelf. Whereas for me with ChatGPT, it's practically every day, and maybe sometimes I need to work a lot, so I spend like 3 hours, or some days it's almost nothing, but it's there, persistent in my life.
How do you discover now if something is a real trend or a fake one within the company and in general? Same principle, like, is there Something real, deep, and lasting? How did you arrive at that philosophy or mindset? Well, I observed a lot of startups. There are many amazing things about working at YC, but just the amount of data you receive—as long as you 're willing to dedicate the time to analyze it and make sense of it—really allows you to uncover a lot of things.
Are there any significant mindsets you used to have that you think are incorrect in the current environment and will become even more misguided over time as intelligence explodes? I mean, I think a lot of them weren't necessarily wrong at the time, although I'm sure I was wrong about many things as well, but a large portion of them are now.
Um, I think a startup today still mostly looks like a startup from 10 years ago because that's what conventional wisdom says you're supposed to do. And sure, it's different in some ways. People say, well, I'm going to hire fewer people and spend more money on tokens for Codex or whatever. But it should probably look very different.
And I've met some people They're trying to create a startup in a completely different way. But most just use it to mean using more models, and that does n't seem to be enough.
Comments 0