Sam Altman on Building OpenAI & Betting on the Impossible

I just brought up Tobi Lütke and the fact that I recorded with him previously. Why did you say you think he's one of the most interesting CEOs right now?

One of the things that struck me most about Tobi is that in the very early days of AI, and then at every moment along the curve as it's developed, he has been the most forward-leaning CEO. He's in there writing the software himself. He's experimenting with it. He sends us extremely detailed feedback on the product offering and on the capabilities of the models. Before anybody else was saying this, he was saying, "We are not an NPC company, and thus we are going to adopt agents. Otherwise we're totally screwed. We're going to build it ourselves." Every time I talk to him, he is at the edge of what anyone, CEO or not, is doing. He builds himself. He has a great, deep feel, and he is always six, eight months ahead of any other CEO.

Do you remember when he wrote, I think this was probably a year and a half ago, maybe 2024, that letter saying, "The first thing you have to do is see if AI can solve your problem"? Even back then, maybe 18, 24 months ago, people went crazy. They thought it was ridiculous.

They thought it was ridiculous. This is my point. He's just consistently been ahead. He has been correct. He's leaned in. He's very no-bullshit, so there's no hype. There's nothing other than, "Here's what it can really do right now. Here's what I think it'll be able to do soon. Here's how I'm going to push the company here." And just an extremely deep understanding of where it's at.

I never even thought about how much of a benefit it must be for somebody in your position to have somebody like that giving you intense, very direct, clear product feedback.

A lot of people send product feedback. He is the only person at the intersection of CEO of a large company and extremely accurate, detailed, cutting-edge feedback.

He told me, I don't know if it was in the episode or after, and he was very adamant: "We're going to look back on 2026..." And I actually want your opinion on this; I don't think I've talked to you about it. "We're going to look back on 2026 as the year that every business was up for grabs." He said somebody was going to build the AI-native version of Shopify, "and it's going to be me." So at night he is clearly trying to rebuild: if you started from scratch, what would you do with the current technology?

That was the other thing I was going to say about him: he does it himself. He's using these tools himself. He's writing software himself. He's trying the models himself. He's trying to reimagine his workflows himself. Most CEOs, when you get to that level, have teams of people managing teams of people who are trying to implement the thing, and they're trying to make you happy, and they're trying to smooth the rough edges. I think it's very hard to get the feel if you're not actually doing the thing, and he does it so hands-on, all night long, as far as I can tell.

I'm not sure 2026 will be the year that every business feels up for grabs. I might disagree with him a little bit there. But I get the spirit of it, and I do understand that it feels like that's happening.

Do you think that's even possible, whether it's 2026 or 2046?

Obviously not literally every business. I think there are some things that are very anti-AI. The better AI gets, the harder, I think, some businesses that have nothing to do with AI will be to compete with, because we'll really want these authentic, non-technological experiences, or we'll care more about sports teams or whatever. So no, not everything. But I think there will be many software businesses that are very up for grabs.

So you disagree on the timeline? What do you mean?

Yeah, I disagree on the timeline. I think it's going to take a bit longer.

Can you say more about that?

I love startups. I think startups are the coolest thing in the economy, and I've spent my career trying to really understand them. And I thought when we got to GPT-4, which was back in 2023, I think, that very quickly after that there was going to be much more disruption, with software businesses up for grabs right away, than turned out to be the case. I think I was wrong about a few things in terms of the speed, but one of them is that the economy just has so much inertia. People keep doing the same things they're doing. They keep buying from the same company. They keep wanting to use their tools in the same way.

I think this is actually a positive in many ways, and it's going to make the big transition in front of us go smoother and slower, and I'm grateful for it. But I think it means we've all been too ambitious on timelines, even with this incredible technology. I think AI is one of the most incredible technologies humanity has ever invented. Society and the economy will adapt more slowly.

It's funny. We were talking before we started recording about all the parallels to history. Obviously, I read history for a living. When you were just talking, I wasn't even thinking about OpenAI and AI and Sam Altman. I was thinking about reading the biography of Larry Ellison in the '80s. He was saying, "Guys, this isn't a software problem, it's a people problem. We have to convince them. We can install the software, and they're not using it. We have to change their behavior. The technology is there. Now we have to adapt the humans so they actually start using the technology."

My own example of this: after Netflix came out and started shipping DVDs, even before they started streaming, it was amazing to me that people still went to Blockbuster. It was incredible to me. I would just watch this, because I drove by a Blockbuster on my way to and from school, and it amazed me that people still did it. That example has stuck in my head, of force of habit and the way people do things. Changing behavior is just much harder than the tech nerds realize.

The missing iPhone moment

So if we go back to the uproar over Tobi writing that letter to the people inside his company: you're adopting this faster than anybody else, because you're partially inventing it. Is there something where you're actually shocked at your own behavior? Where you're like, "I know there's a better way to do this. I'm even creating the product that could make it better, and yet I still can't get over this force of habit"?

A hundred percent. I love that. I don't think anyone's ever asked me this before. I've been waiting for this question.

The thing about myself that feels most psychologically inconsistent is that I have been using computers the same way for 20 years. I now have a magic thing called Codex, so do you, so does everybody, which means I should be using my computer in a completely different way. I should not be clicking around, pasting from one messaging app to another. I should not be scrolling mindlessly through my emails, trying to figure out which one is least painful to open and respond to when I don't want to be dealing with it. I should not be keeping a to-do list and doing these rote computer tasks the same way I have for so long.

And yet there's something encoded in my mind that says doing this kind of stuff is what it means to work, what it means to be productive. If you asked me, I would never say I like doing it that way. In fact, I would say the opposite, and I think I would mean it. But by revealed preference... I have a better way to do it now. I can do it faster. I could be using Codex for more of my day-to-day: got to get through this stack of emails, got to do this stuff on my to-do list, got to deal with all these things. And I still do it the old way, and it makes no sense, other than that I must secretly like it or feel good about it.

What do you think has to change for you to actually adopt your own product in a deeper way?

I don't really know. It's happening gradually, and this might be the right answer: these things have to happen gradually, and totally changing someone's ingrained habits and workflows is difficult. I think there are better products we can build with this technology that will make it more seamless. But right now it feels like we're all straddling two worlds. We still have a computer we can use the old way, and we have Codex that can use our computer in this amazing new way, and we're not sure which to use when, for what. I think this is mostly a product failure.

The phase we're in now reminds me of smartphones before the iPhone. I was an early adopter. I had a Palm Treo in 2003 or '04, whenever it was.

Did you ever have a Sidekick? Do you remember those?

I never had a Sidekick. I thought they were super cool. I wanted one. And a lot of the technology was there. It was missing multi-touch, but mostly it was missing the product ideas that made the iPhone the iPhone. I feel like we are now in a world where we have all the technological pieces, but we haven't had the iPhone moment of completely changing how someone interfaces with technology.

Research and compute

We were talking about Tobi, who's out there building these things himself. We have a mutual friend in Josh Kushner, and he says there's a big comparison to be made between the way Steve Jobs thought and ran his company and the way he thinks you do. Jobs obviously wasn't the one writing the code. He wasn't building the hardware. But it was, "I am patient zero. I am making products that I myself want to use." And essentially everything we saw with Apple was basically what he wanted.

There's a great story in one of the books where they were supposed to have a meeting on, I think, one of the new MacBook laptops. The team prepares this huge presentation for Steve, and they're really nervous because of his commanding presence. They think it's going to be an hour-long meeting. He walks in, and he shows them a device: "On, off." You press the button, it comes on. Press it off, immediately. And then he opens up the MacBook, and there's a delay. He goes, "Make this," meaning the MacBook, "like that." And then walks out of the room. That's the whole meeting. There are a lot of examples like that in the history of Apple. How do you approach it? How do you improve the product? Are you just doing it through your own needs?

Most of my effort right now is on research and compute. I would love to be able to spend more time on product. We have great people thinking about product here, but the most important thing we can do is create smart models and be able to run them efficiently and abundantly for a lot of people. If we can get that right, I believe everything else will follow. Philosophically, I'm very inclined to try to find the high-leverage, difficult problem that will continue the exponential. And for us, that's models and compute. I also just think those are problems that naturally suit me.

Why do they naturally suit you?

To scale compute in the way we're doing, it's a complex supply chain. There are a lot of interesting partnerships to figure out, which I like doing. There are interesting financial challenges in how you're going to finance what is probably already, or at least is rapidly becoming, the most expensive infrastructure project in history. And the technology questions that go into building out compute at this scale all come together: from designing your own chip, to the supply chain of fabs and the people who make racks, to the power systems for these things. I've always been interested in energy. So there are a lot of interesting problems across technology, business, policy, supply chain, and logistics, all together, in building compute at this kind of scale.

I used to be a startup investor, and the thing in my career I've found closest to startup investing is managing a research program. There are all these ways in which they're really different, too. The average researcher and the average founder look different on the surface, for obvious reasons. But there are a lot of similarities in how you find the non-consensus bets, how you decide where to have conviction, how you understand what exponential growth looks like, how you manage outlier talent, and even more, how you identify it. This is the research building that we're in, and it's where I sit.

Power laws

Say more about the parallels between what you learned in startup investing and doing research.

One big one is the power law. People talk about this all the time in investing: you have to reprogram your brain, because we don't seem naturally well suited to think this way, where your best investment will outperform all your other investments put together, and your second-best investment will outperform everything after it put together. And AI research, at least, is like that as well. When we started, people thought it was totally unlikely, almost impossible, that AGI was possible.

What year is this?

2015. We just got hammered by all the intellectual giants of the field for saying we were going after AGI. And then when we started really focusing on large language models, we got hammered again: "This is completely ridiculous." And we understood, or I understood at least, from my startup background, and I think other people understood it in other ways, that high-risk bets are okay as long as you take the ones where, if they work, it's super valuable. Research looks this way. And the kind of people who make great researchers are non-consensus, fresh-approach, high-energy, nonstandard is the word that keeps coming to mind, people.

You've got to say more about nonstandard. Can you be more specific? Are they spiky?

You don't want to fund a founder who has a very slightly different take on the same idea as the last thousand people you talked to, who has tried to convince you, and maybe themselves, that they're somehow completely different and doing something totally new, but who is mostly trying to fit in with the herd, be on the same track as everybody else, and do what they're supposed to do, which is start a startup. They've heard Peter Thiel say enough times that you're supposed to be doing something different that they try to emulate it, but they don't really mean it. It's very clear to me when you have someone who just thinks differently from most other people and is willing to stand by convictions that are very unpopular, and may well be wrong, but if they're right, they're going to be really right, versus someone who is a thin veneer on the same idea everybody else has.

In late 2015, when we were starting OpenAI, there were very few AGI efforts in the world. There was DeepMind, and one or two others I can think of. It was a very non-consensus thing to do. In that same year, and I'm picking on it just because it came to mind, there are other categories too, there were probably many, many thousands of founders starting photo-sharing apps. That was probably not as good a thing to do. Today, a lot of people want to start AI labs. There's some handful of people, two, three, whatever, doing something completely new that actually wasn't possible until AI got this good, but doesn't seem like a good idea yet. And that's the thing that, as a startup funder, I always wanted to fund, and the thing that mostly worked for me.

There's a similar thing with researchers. There were a lot of researchers who would chase whatever the last thing was that worked, and there was a small number of researchers who had high conviction in a new idea. And I think we were, and are, the best research lab for those people.

I remember talking to Demis about this a few months ago. His thought was that in terms of chasing after it the way you guys are, there are the big three players, and with the money, the capital required, there's not going to be a fourth bigger player. But he said there's some, I forget the number, I'll just make it up, say 10 percent chance that some monk researcher approaches it from an angle we've never even considered.

Totally. I don't know how to put a number on it, but there is some chance of that, for sure, and I love that. I think that's why stuff stays exciting.

From AI-obsessed kid to investor and back

Okay, so what's confusing to me: you went from founder to investor and back to founder. But why in 2015? What got you interested in artificial intelligence to begin with, to the point where you're saying, "Hey, this is such a non-consensus thing. People think I'm fucking crazy. I'm going to do it anyway"?

I had been interested in AI my whole life. I was a very nerdy kid. I was the kind of kid who spent Friday nights playing on my computer and watching sci-fi, reading sci-fi. And I always thought AI would be the most amazing, craziest thing. I never thought I would actually get to work on it, but I always loved it. I even came to college partly to study it. I worked in the AI lab the summer between my freshman and sophomore years, and nothing was working. In fact, very memorably, a professor told me, "You could try all of these things. There are all these directions. The one thing we know doesn't work is deep learning. We tried that for a long time. It's the most guaranteed way to have a bad career." And I was an impressionable college freshman or whatever, so I assumed that was true and pursued these other things. It was clear to me at the time, and this is now 2005, that AI was not working.

And I happened to accidentally get into startups, and then very much fell in love with them, so I wouldn't even call it a career detour, because looking back it was super helpful, and becoming a startup investor was great. The normal career path in Silicon Valley, or not the normal, a common path, is that you're a founder, and then you semi-retire and become an investor.

We don't want that. I heard you say you're going to work on this for the rest of your career. Hopefully you're going to come on the show multiple times. I'm going to hold you to this. We don't need more founders who retire and invest. We have enough investors.

No, I don't want to. We have too many investors.

Thank you. I want to talk to you about this.

But what I was going to say is that the fact that I got to go in the other direction, investor first and then running a company, is pretty unusual.

Very unusual.

And I'm super grateful for it, because if you really study and watch companies as an investor, you get this unbelievable set of learnings and pattern matching, which has been super helpful to me running OpenAI. But it's the opposite of the normal direction, so it's a very rare thing, and I strongly recommend it.

Why is it helpful?

If you're running a company, you've faced some number of similar decisions in your past, some number of crux decisions, and you've seen what works and what doesn't. If you have to make a high-stakes strategy shift, or fire an executive in a really messy way, you have whatever your own limited experience was over the last five or ten years you've been doing it. But as an investor, you watch all the crux moments. You don't get the operating practice you get day in and day out running a company, but you've seen a lot of the big crux moments. A lot. You see those all day long. So the wealth of the dataset I had there was awesome.

So that plays in your head when you have a decision to make?

Yeah. I'm like, "Oh, this is what happened when this company had a similar thing," or, "I saw this founder make this mistake," or, "This founder got it really right."

We were talking about how you studied the Industrial Revolution, and about some great biographies we've both read. My friend Daniel Ek says this about me, because I think that's the benefit of doing this project, my other podcast, Founders, for ten years. He says, "You're like an LLM trained on history's greatest entrepreneurs, but with the temperature turned up, because you're fucking crazy." Because I'm super passionate about it; it's a weird thing to be obsessed with dead entrepreneurs. But it is helpful. I'll be talking to a founder, and they'll talk about something they dealt with, and I'm like, "Oh, well, Carnegie did this, and Rockefeller did this, and you might want to try this."

And do you find you have one big insight from all of that, or is it that for any given scenario you have what all these people did and how it comes together?

I think it depends on the personality of the founder. When I was reading... I've read your blog for years. I think you're a great writer. You're very succinct.

Thank you.

The brevity is really appealing to me, and I love numbered lists. It's weird that we both write the same way. I'm reading your blog, and I'm like, "This is the exact conclusion I would come to based on all the reading." There's this handful of principles that could be applied, but it really depends on who the founder is and what they want to do.

This is what I'm trying to understand. Let's go back to what we were just talking about. You're making a big jump, because from what I hear, you're one of the best investors of all time in Silicon Valley. You could just be rich and not really have to work, because investors are kind of lazy. I'm just kidding, by the way. Kind of not.

Having done both, I think I can say it's much, much, much harder to run a company than to be an investor.

Exactly, and that's what people should be doing, in my opinion.

I agree with you.

So then you're like, "Fuck that. I'm not going to take the easy route. I'm going to do the hardest thing ever, the thing people think is impossible, the thing you're made fun of for." I still need to understand. So you're into it as a kid. Why would it appeal to a kid? You were living in St. Louis at the time?

Yeah, I was living in St. Louis.

Why would AI appeal to you back then?

I think it appealed to every computer nerd. I don't think it's that unusual about me. It just felt impossible. I think most people would say, "Of course that would be the coolest thing ever, but it's totally impossible." I think the weird thing about me was, "Okay, let's try." But I think everybody would think it's awesome and something to go for.

So wait, that was a personality trait of yours as a kid? When people told you that you couldn't do something, your initial response was resistance?

Not resistance, but: "Are you sure? Why not? Let's try. Let's see what happens. Maybe I can. Maybe we can." I was a very optimistic kid, and the more impossible something seemed, the more intrigued I was. The idea that we could invent a technology that would let us do everything else, that would empower people in a way no other single technology could, always seemed innately, incredibly appealing to me. I want that thing. I want to be able to do everything else.

Another thing that's been a personality trait for as long as I can remember is that it's interesting to really give people a lot more power, a lot more ability. In some sense this is the whole arc of technology, and I was for sure always a technology nerd. But AI is the strongest version of that I can imagine.

What did you think it would enable back then? When you were a kid: "This seems like a cool technology. I want to do X. I can't do X unless AI is invented."

It's always hard to remember how much of this is stuff I actually thought at the time versus how much...

Versus what you're trying to build right now.

Yeah, how much my current work has colored my memories of it. For sure, as a kid I was very into robots. We had a robotics club at my school, and the robots at the time were laughable. I even remember at summer camp we had this little turtle you could control with a computer, on the floor or on the table, and I thought that was just the coolest thing. There's something about physical stuff moving, controlled by a computer, that I always thought was amazing.

Impossible problems

Now I'm extremely interested in what AI can do to advance scientific discovery. In my memory as an adult, I think I thought that was cool as a kid too, but it feels implausible, and I assume that's an example of where the memories have gotten colored. But now, the fact that we can have AI go discover new physics and cure diseases, and what it's already doing for math: I think this will be one of the most important areas, even more important than the automation of other tasks. AI can just help us understand more things. We were talking earlier about the book The Beginning of Infinity, and rereading that book from today's vantage point, I'm like, "Man, AI is really going to help us do this important thing of understanding everything, or as much as we can."

I was definitely interested in the Star Trek version of huge prosperity and abundance, and what AI could do to drive that. Maybe the memory of being interested in science is more real. I also loved science, and this idea that because we were smart, we could figure out how to understand the world and make predictions and do things we couldn't do without that deep understanding. I don't know. That seems innately awesome.

It's interesting how consistent over time what humans want from AI has been, because what you're describing is very similar. I just reread the biography of Claude Shannon for the second time, and I had forgotten, because I hadn't read the book in maybe five years, that he and Alan Turing used to meet every day for coffee when they were both at Bell Labs. This is the 1940s. And they would just talk about AI. They were both obsessed. They thought it was inevitable back then, and they thought it was going to happen in about 15 years, so around 1955: that we were going to have computers, which didn't exist yet, they had the analog versions, that were smarter than humans. And anybody who thought that wasn't going to happen, they thought was absolutely ridiculous. And someone asks, "Well, what would you want the computer to do?" He's like, "Solve math problems, write poetry, cure diseases." You hear this over and over again.

I've read a bunch of things those guys wrote at the time, and I'm so sad they're not here to see it, because they were so right about everything. We're finally at the moment where AI is solving novel math problems. It's discovering other stuff. And you can argue about how good it is or not, I would say not very good, but it's writing poetry. It's here. And I think those guys would have said, "All right, you've done it. This is it. We've got it." And that would have been so cool.

This is the weird thing, where everybody's just like, "Oh, it'll never do X." I talk to people in the music industry: "It's never going to make great music." And then they ask, "Well, do you think it's going to make a podcast?" And I say, "Of course it's going to. It's going to do everything we can do, at least as well, and I would say even now, in some cases, better." It's a very bizarre thing, this idea that it will never surpass what's happening at the current point in time that I happen to be alive.

Human connection

There's a deep human psychological flaw there. But here's what I think is a more interesting question. Let's say it does make a great podcast. Two AIs are having a more interesting conversation than you and I are. Do you think people will care, or will they want the one with the real people, because we're all obsessed with people, and the fact that it's not real people matters?

For this, the human version is more interesting. It's, "Oh, these two people that I may be predisposed to like or dislike are having a conversation that's interesting to me." Maybe my other podcast, where I'm just saying, "Hey, here are some interesting ideas I read in this book," could get disrupted in some cases, but especially for people who were born before this happened... Maybe it's different for your son.

Maybe.

But for me, I think humans are always going to be drawn to humans.

I really deeply believe that. I think there are a lot of other jobs that could face significant transition, but stuff that's about people, about people's connection to people and people liking other people, feels like it gets more valuable in a post-AI world, not less.

I may actually be the wrong person to talk about this, though, because even though my entire work is digital and broadcast all over the world, I deeply desire more of an analog life. I like reading physical books.

I do too.

When I was talking to Kelly, I was like, "I don't want to get on Zoom. Call me, or we'll talk in person." I like physical shit.

That's interesting. I'm like that too. I don't read e-books. I don't like Zoom meetings. I like to be with people in the real world.

I definitely think there's a subset of weirdos, and there are probably a lot of them living in this city, who don't like humans and only want to communicate with computers. But I think that's a tiny percentage of humanity.

I think it's a tiny percentage of humanity. This is why I think the world is, on the whole, not going to be that different, even with superintelligence. People are still going to be very fundamentally wired to care about other people, to want to be around other people, to interact with other people. There will be some people who get obsessed with the models and think humans are in the way, or a danger to be contended with, or whatever. But for most people, people will be the whole point.

Two risks

I think it's very important that when we do find people like that, they get called out, and we make sure they don't acquire power.

I certainly agree with that. Maybe the two big risks I'm most worried about with AI, which are a bit in tension: one is loss of control, where AI somehow becomes too powerful in a way that we can't guarantee control of it. And the other is power getting too centralized, where you have one company or model or person with too much power. In both of these, the fundamental thing is that I think it's a very anti-human position for either to happen. The right approach is to say: we want people deeply in control of the future. We want people deeply empowered. People are the whole point of all this. We are not going to sit here and gradually hand over control to an AI model because we don't trust or like people. It's a very misanthropic thing to say we're going to put all our trust in this model and let it have all the power and decision-making over the world. But I think there are some people in the world who think that's the right outcome.

There's another version of this, which is: because we don't trust people, we have to limit who gets access to this technology and how they can use it, because all these terrible things could happen. And out of fear of those, we're going to concentrate power in the hands of a few companies, and we're not going to let other people use it, but we'll give them some benefits. My caricature of this is that there are some people in the AI field who effectively say, "We're going to give the world a cure for all diseases, and we're going to make stuff really cheap, in exchange for people giving up their autonomy and their impact on the future, and their power," also in the name of safety, and also with just absolutely rampant inequality. There will be people with access to huge amounts of wealth and power, and everyone else just gets a pretty good everything. This is a terrible sales pitch. It's a very anti-human sales pitch that somehow people feel willing to make.

Why do you think they feel willing to make it?

I think it's fear, and power. When people talk about the risks of AI, I think there are a lot of people who are so nervous about the magnitude of those risks, and so taken by it, and who feel such a need to protect the world from it, that they say, "We should trade off a lot of liberty for safety here, because this is unlike other risks we've seen." But then I think that also ends up being a way to justify a lot of power-seeking behavior.

Learning from reality

Everything I read, like what I was just saying about Claude Shannon and Alan Turing, at least in the books I've read, is more optimistic: "We're going to invent things that make our lives better and can do things for us."

You're totally right that if you go back to the Claude Shannon, Alan Turing era, they talked about how wonderful AGI would be and all the things it would do. And when we started, we really had a lot of pressure from the doomers. Now, the part of the doomers I agree with is that this is a powerful technology, we should err on the side of safety, and we should act with caution at each level of the technology. The part I don't agree with is that it's an unsolvable problem.

If you go back to the beginning of OpenAI, I think there were two widely held opinions. Number one: not at all, and certainly not in ten years, were we going to build something very AGI-like. And then, conditioned on doing so, we certainly were not going to be able to make it safe. If you had an AI that was smarter in many ways than most of the smartest people, the doomers would say, surely at that point the world would have been destroyed. The alignment thing would have failed. There were these very confidently held positions about what would happen.

A decade on, we have built something that I think most people at the time would have said is very AGI-like. A lot of good things have happened, and the crazy bad predictions of the world ending have not. So I think that should update people's predictions about the future. There are still higher-stakes challenges in front of us to solve. But our approach, and this is another thing I learned from startups, is that the way you do things is to put things out into the world, get feedback from real customers, and see where they break and where they don't. That's the way you make a good product. It's also the way you make a safe product. And we've made way more progress on AI safety than I think most people thought we would when we started.

Why? Because there are a billion people using your products on a weekly basis?

And each time we get a new level of model, we put it out in the world and see what works and what doesn't: where people need us to relax the guardrails because they have good things they want to use it for, where we have alignment failures, where we have safety-system failures. ChatGPT has been out less than four years. A billion people use it, for sensitive, important stuff. And the fact that we can deliver something that's broadly considered safe (of course there are issues with it) in that short a timeframe, with such a powerful technology: I think there's no way we could have done that in an ivory tower. This is how I believe you build good, safe, robust, useful technology and products, and I think it's a great learning from Y Combinator. It would have seemed totally impossible to most AI safety people to get to this stage and still have the level of safety guarantees we have. Now, I do think it gets harder from here. But I don't think you're going to solve it by disconnecting yourself from reality.

Why does it get harder from here? Because the smartest people in the world are about as smart as the smartest models in the world, and that's about to flip?

Directionally, I think that's right. The models are so incredibly capable, and improving on such a steep trajectory, that the unknown unknowns... Maybe they don't get harder relatively, but from an absolute perspective they seem harder. And I think we'll have to make a bunch of difficult decisions about when we delay development, when we say, "Okay, you know what? Let's have contact with reality now," or, "Let's wait longer and really study this more."

I was talking to someone recently, and something that stuck in my mind is that the FAA has helped make flying incredibly safe. On the surface, flying seems like an extremely dangerous thing, and you probably get on an airplane without giving it much thought. Airplanes are not that old in the long trajectory of human history, and this was certainly not the case at the beginning. They have extremely robust accident reporting, extremely clear-eyed. They never try to hand-wave over something. They want to extract as much information as possible. And in some sense, with any new technology, I think an approach like that works very well and is often underappreciated.

So when we started deploying our models, when we said, "We're going to put ChatGPT out in the world," we knew the model was imperfect. We knew it hallucinated. We knew it could do these other things. But we also knew the world had to experience this technology, that we had to learn how to make it safe, and that we had to put the power in people's hands. We cannot just use this to impose our worldview. We cannot go sit in a lab and try to think through all the impacts, which won't work anyway, because society and the models are going to co-evolve. We all have to do this together, as this joint project, and then we'll do very good accident reporting. We'll study it when something goes wrong. We'll put out a very clear postmortem. We'll learn as much as we can. We'll not only improve our own technology and products, but try to share those learnings with other people building AI. I think that's worked surprisingly well so far, and there were good examples of it from the history of technology, and good examples from startups.

Why people fear AI

One thing has to be disorienting for you, because it doesn't make any sense to me. Essentially, what I focus on is entrepreneurs and entrepreneurship. All the podcasts I make are for the benefit of entrepreneurs. I'm glad other people listen, but it's heavily focused on trying to serve them, trying to find useful information, whether it's in a biography of a dead entrepreneur or in talking to somebody like you, that other entrepreneurs can benefit from. And entrepreneurs love AI. But I'm trying to figure out if you can help me reconcile this: everybody uses AI, and everybody hates AI. What the hell is going on there?

Well, people are always afraid of rapid socioeconomic change. We talked about the Industrial Revolution earlier. I love reading about previous technological revolutions too, and people did not have universally warm and fuzzy feelings about the change happening throughout the Industrial Revolution. I think that's probably a good feature of human society, that we have some built-in inertia, some skepticism of rapid change. It probably helps society in times of turmoil, or in times of localized craziness or whatever. So some of it is probably good, and it's a feature of human biology, which I believe in never trying to fight too hard.

I also think a lot of the people building AI have been out there saying, "There's a 25 percent chance we're going to destroy the world, and yeah, we're going to race ahead to do it, because otherwise the bad guys will do it first." Or, "Man, this thing is going to be really terrible, and 50 percent of the jobs are going to go away in the next year, and we hope you're all okay, but it seems really scary." We have not, as a field, done a very good job of explaining to people what the benefits are and how the downsides can be mitigated. And we certainly haven't done a good job even when people have had answers, like saying, "Well, there's going to be universal basic income," or, "Work will be optional," or whatever.

There's been very little discussion of how and why it's important that people have more power and personal freedom in the world, not less. And that matters a lot to most people. The ability of people to influence their own future, and collectively to design where society is going to go, and the autonomy that comes with that: very important. I don't think a lot of people in the AI field spend much time thinking about it, reflecting on it, acknowledging how important it is to other people, even though they feel it for themselves. So to go back to that characterization of the sales pitch from earlier, I don't even know if "characterization" is a word...

It is.

"Dear peasants, we will bequeath upon you these gifts: a cure for cancer and material wealth and some things and great entertainment. You stop complaining, and we'll make all the decisions about the future, and just trust us. We'll be benevolent dictators." Not good. Not good.

The small-business boom

As a lover of entrepreneurs, and as a student of what has made the incredible economic miracle of recent centuries work: really empowering people to go do new stuff, to push on the things they believe in, to have the freedom to create companies and invent technology and pursue ideas, and the system around that which makes it happen. There is nothing I believe in more strongly. And I think even people who don't see themselves as entrepreneurs, who may never want to start a big company, understand how important that is. So when they hear people implicitly or explicitly saying there's going to be less of that with AI, because a small number of people are going to have the power, but they're going to make great decisions and keep everybody safe, I think that's very scary to them.

I also think that even if most people don't want to start really big companies, a lot of people want to start smaller companies, and that has been hard. It has required a fair amount of privilege and luck and resources. And we are about to see the greatest boom in people starting smaller businesses that we've ever seen. I think AI is empowering that. For some reason the field, including us, hasn't talked about that enough, even though we see all these signs of it and it's great, and we haven't built enough products to accelerate it. But I think we're going to see a lot more of it.

It's funny. We talked about Tobi Lütke earlier. At the end of our conversation, and I think it's in the episode, he mentioned something I hadn't thought about. He said, "Oh, yeah, you and I are in the same business. We're both trying to create more entrepreneurs." He's building infrastructure for entrepreneurs; I'm building educational and inspirational podcasts for them.

And the crazy thing about what you just said is that, out of any new industry, the AI industry is doing the worst job of this I've probably ever seen. I think part of it is just that you guys have to get out there and talk about the stuff you're seeing, and educate. There's a great book called The Intel Trinity, about the story of Intel. It's called "The Trinity" because the three main players are Bob Noyce, Andy Grove, and Gordon Moore. There's a great story in the book that I've never forgotten. They go from inventing, I think, the integrated circuit to the microprocessor, and they realize the technology is so important that it would scare their potential customers. So those three people stopped what they were doing and went out and started educating potential customers, investors, the entire country. At one point they were putting on more classes than the local community college had in its entire course catalog. That's how much of a top priority they made it: "We're going to get out and educate about this new technology." So why isn't anybody in AI doing that?

No excuses. We should be doing more. I think we've tried versions of this. We haven't gotten it quite right.

Well, you're doing it right now. This is one of the reasons I wanted to talk to you. I use AI all the time. I think it's fascinating. But you have a view that makes mine look like the view of an ant. There's so much in your head that I want to get out: "Hey, you have all this context." And you're inventing this incredible technology. I would love to know how other people are using it.

Context and memory

Actually, before we even get to other people, I heard you say something interesting that ties into what you just said about how we're inventing technology, the technology's improving rapidly, but adoption should be slow and deliberate. I heard you on another podcast say, "Hey, I'm even considering how much I should let AI see of everything on my computer." Do you want to talk about that?

With the latest generation of models, I don't want to say they feel smart enough, because I think we should always aspire for them to get smarter, but they're pretty smart. And at this point I feel more limited by the amount of useful context the AI has on me. I want the AI to know as much as it can to help me. I want it to be doing things I can't or don't want to do on my own. I'm not going to read every post on our internal Slack. I'm not going to read every story a customer has about where ChatGPT worked for them or failed them. I can't. And then there's other stuff: I probably could read more research papers than I do, but it takes a lot of mental energy. But I would love to have an AI agent that is constantly trying to be helpful to me, that can look at and understand more context than I can, or have time or energy for on my own, and can bring that context to bear and give me good advice when I have to make a decision.

So I think we've focused, correctly, so much on model intelligence that on the product side we haven't yet thought enough about what it means to give a model more context than any person could have, and to help advise that person on their big decisions. My sense is that we're just on the precipice of a very different way of working with AI, on an axis where people just can't get this good. There are plenty of very smart people, but there's no one who can read tens of thousands of pages of context in a few seconds and really use it accurately. That's something AI can do that is going to be very new, and an incredible supplement. You've read all these biographies. There are probably times you vaguely remember something where, if you could recall the specific anecdote, it would really help a particular entrepreneur in that moment, when you were talking to them. But maybe you forgot it, or you don't remember it exactly right.

I built my own AI tool.

There you go.

I use it internally. You know what it's trained on? Since 2018 I've kept every single note and highlight from every book I've ever read in this database, and I used to search it. And then when the work you guys do came out, I trained it on that: every note, every highlight, and then all the transcripts of my Founders episodes. I use this thing every single day.

That's awesome.

To make every single episode. So I'm working on Claude Shannon, say. The Claude Shannon episode came out, I don't know, two or three weeks ago, and I'm asking it questions: "Hey, what did Bob Noyce say about this?" Or, "What did Rockefeller do about this?" I made that episode. I read the book. I took the note. I don't remember it, because it was seven years ago. It's incredible. This is what I mean. It's fucking awesome.

That is so cool.

A platform, not a product company

Let's get into how you think about running the company. You said your main focus is getting more compute, and then research. So you want the models to be the best in the world. But how do you think about whether you have to build your own products? You built Codex. I don't even know the fucking product lines. Where's all the revenue coming from?

Actually, I think we should be more of a platform company than a product company. We'll build products, of course, but the products...

Well, how many products do you have? Let's back up. How many products do you have now?

We just merged ChatGPT and Codex together. We used to have ChatGPT, Codex, and the API. And Codex, unfortunately named, was not just for coding. It could do any kind of work, which confused people.

I'm confused by that.

Yes. As were many other people.

What I think most people want is a single interface to their own personal, or their company's, AGI that can help them with whatever they need, and then the ability, with an API, to build anything they want on top of it. That's the platform we should offer the world. We're going to sell great AI. At every point on the cost-performance curve, we will be the best. If you want really high-end AI to discover science, that's great. If you want really inexpensive AI to do a massive volume of work that maybe doesn't require genius-level intelligence, we've got you covered there too. People talk about this in different ways, a new utility, a new commodity, whatever you want to call it. People want to use a lot of AI, they want it at low cost, and they want it to be fast, to work well, to have their context, to be smooth. We've got you.

And then there's a single product, which is: I need to ask the AI something. Eventually maybe it's that the AI should proactively offer me things. But you'll have this interface, which started as a chatbot and now also has coding agents, and which I think at some point will feel like a more persistent agent, to this AI running on whatever you need it to run on. But that's it. I don't think we should go build every product category. I don't think we should try to compete with all our customers. I don't think we should try to subsume the entire economy. I think we should offer this platform and try to have a hundred million new businesses and eight billion people use it in all kinds of new ways. So: one direct interface to the product, and one API for people to use however they want. Those eventually come more and more together too, and then it's all about what people do with it and build on top of it.

Killing good ideas

What mistakes did you make to have to learn that? I feel like you've had to kill some good ideas in order to go after the great ones with your full intensity and focus.

Yes. I think killing the good ideas, sacrificing the good ideas to go after the great ideas, is the hardest lesson for any entrepreneur or business to learn. It sucks to kill good ideas. And no matter how much you think you're going to do it, everyone, maybe just by the nature of who chooses to be an entrepreneur, seems to do terribly at this. I'm terrible at this. I know I'm bad at this.

But last year, for example, we killed Sora, which was a good product, fun and cool, but it used a lot of compute and wasn't as important as Codex, which is where we put the compute. We killed our web browser, Atlas, which again I think was a great product. I think it was the best web browser, but it wasn't as important for us to focus on as somewhere else we could put that talent. In a world of limited compute, limited people, limited resources, we thought really hard, and we said, "You know what? General intelligence for knowledge work, and eventually for science, is the most important thing we can do." Anything that goes into that, upstream of generating that intelligence, building our own chip, building our own data centers, writing good infrastructure software, and obviously training models: that's all really important. But then let's just offer this AI as a service and get people to use it for intellectual pursuits, for work, for scientific discovery, to be more productive in their personal lives. Let's have a flexible, general platform, and not do a lot of other things.

Peter Thiel and Paul Graham

Anybody engaged in complicated work, and you've got to be toward the top of the list of anybody alive right now, needs somebody to help organize their thoughts. It's extremely beneficial. You see this in every single biography, in every history. You need somebody to talk to. There's actually a funny story about how extreme this can be. Charlie Munger has this thing called the orangutan theory. Have you ever heard of it?

No.

He said a relatively smart human could go in, sit down with an orangutan, tell it all his problems, tell it everything that's on his mind, and the orangutan obviously says nothing. The human leaves, and the human's better off. Just from being forced to put your thoughts into some kind of structure. Now, obviously, it's better with a very intelligent partner. Munger played this role for Buffett. Buffett's one of the most intelligent people who ever lived, the greatest investor of all time, and he still needed to organize his thoughts with somebody else. You have almost a singular lived experience, especially for somebody as young as you are. So I'm curious: who plays this role in your life? Who do you go to who can even remotely empathize with what the hell you're dealing with on a day-to-day basis?

There are three categories here. One: a lot of the researchers who have been here forever. We've all been through it together, and we've developed a set of shared language, intuition, standards, whatever you want to call it, that I haven't been able to replicate with anybody outside the company, when it comes to the shape of what's happening, what might happen next, and where the technology is likely to go.

On questions of business and the world, for a long time in my career Paul Graham and Peter Thiel have been two of the people I've learned the most from, across lots of different phases of my career, and they're still the two people I go to if I have a very non-obvious problem I'm stuck on. And there I haven't found anyone else, after a lot of looking, with the same ability to think in a super nonlinear way. If what LLMs do is predict what word comes next, those are two of the people for whom I can least predict what word is going to come next, and that's a super valuable skill. You go to them with, "Oh, man, I feel really stuck, and I've thought through all these options," and they can tell you, "Eh, I think none of those options are good. Here's this thing that now seems totally obvious and correct that you didn't think of." A completely different view that you haven't heard anywhere else.

Is this more of a prompt for your own thinking, as opposed to explicit advice, "Do X"?

It's often, "Here is a specific thing."

Really? So what would be an example you could share from Peter? Peter's very fascinating to me.

He is very fascinating.

And it's actually who I thought of. The reason I thought of this question just now is that you said, "We had to kill these good ideas for the great. We're cutting Atlas. We're compute-constrained. We have to focus, focus, focus." That's something that's very obvious when you listen to him talk about the importance of focus: if you have something that's working, make it work better and keep going down that line. Taking an hour away from it to explore something else is too expensive. You should just go deeper on what's already working. There's a lot of value at the extremes.

After we launched ChatGPT, it was this weird thing, because people didn't really know what to use it for. It was growing super fast, but it felt very unstable, almost like low-value growth. People were using it just because they were interested in talking to it and seeing what it could do. So there were a lot of people in the company saying, "Ah, we've got to figure out something else. This is not sustainable value." And we were talking to him about a list of five or six other things we could focus on instead. This was maybe two months after ChatGPT launched, something like that.

And he said, "It's an obvious mistake to do anything but this. Beyond the fact that it's growing, which is rare and great..." And it wasn't growing as fast then as it did later. He said, "The power of this is the power of the Google text box. It's a text box you can type anything into, and it does the right thing. And it doesn't match the current Silicon Valley wisdom that you've got to have feeds, you have to have a network effect, you have to have..." We had none of those things, and that's why everyone was worried. "You have to have a way for people to build up more context." This was before we had memory. "Are people going to get locked in?" He said, "All that stuff: people have just been chasing the Google business model for 20 years, and this is the first thing that's come up. And clearly the empty text box worked for Google, so why don't you just double down on that? It's growing, it's very flexible, and it has all the signs, other than that it doesn't fit the current Silicon Valley wisdom." And I was like, "Okay." So we went super hard on ChatGPT, and it was great.

There's a simple genius to what he said.

Sometimes there's more complex genius, but that was an example of very important simple genius.

What about some advice, or guidance, or a direction Paul Graham pushed you in?

When you just said that: this is a meme for many YC founders. You would go see him for office hours, and he would say, "You know what you should do?" And he would shake his finger like this. "You know what you should do? You know what you should do?" Sometimes the thing that came after that was great. Sometimes it was terrible. But the important thing was that there's a creativity, an open landscape, a "Let's try a lot of things."

We talked about the spirit of iterative deployment, and about startups in the same way. He really, I think, pushed the startup ecosystem into this world of: you've got to ship a V1 you're embarrassed of, early, and it doesn't matter if it could be much better. You'll get it much, much better because of the feedback from customers. I don't even think I asked him before we launched ChatGPT, "Hey, do you think we should launch this thing?" But I knew what he would say. I knew it was still early. I knew it was still embarrassing, and I knew the right thing was to get it out and in front of people.

So wait, your mental model of Paul Graham is so complete that you don't even have to call him?

No, just in this one case. That's the one where there's certainty.

Let me tell you something funny. A few months before he died, I went to Charlie Munger's house and had dinner with him, and I asked, "How often do you talk to Buffett?" He goes, "Never." I go, "What?" He goes, "We talked every day for hours and hours. Buffett can just pretend to pick up the phone and call me, and he already knows what I'm going to say." That obviously comes after 65 years of working closely together, but I thought it was hilarious.

That is hilarious. That's a really funny story. No, there are many times I can't predict what he's going to say; that's why I think it's valuable. But on launching when you're embarrassed of the product, I know what he's going to say. That has been, I won't say the most valuable piece of tactical YC advice, but it's up there. Looking back at all my data points on YC founders over the years, I'm astonished at how much moving fast and being iterative correlates with success.

How Y Combinator shaped OpenAI

Okay. You've mentioned YC way too many times in this conversation, so I have to explore this. We talked about it before: listen, I'm not a journalist. I'm an enthusiast. I don't have a list of questions. I have a world-class founder across from me, I want to know what the hell is in this person's mind, and I want to extract information, selfishly, for me. And I'm shocked at how much you reference it in conversation. Going through YC and then running YC, being affiliated with it, has clearly been very impactful on your life. Can you expound on this?

There's some band that wasn't that successful. They didn't sell that many albums, but they influenced all the musicians who came after.

I think it's called The Band. Literally, Rick Rubin told this story.

It might be The Velvet Underground, but you know the idea I'm getting at, whether it's The Band or whatever. It's not fair to talk about YC that way, because YC, measured by traditional metrics, market cap created or whatever, would be one of the handful of most valuable tech companies. But the degree to which YC influenced everything that has happened in the last 20 years of the tech industry, startups, entrepreneurship, whatever, is, I think, only partly understood. OpenAI is an example of that, not just in how we've shipped our products into the world, but in the philosophy of how we run our research lab. And I think if you talked to many of the other people running this generation of large tech companies, they would tell you similar stories, even if they didn't go through YC.

But what's happening there? Is it an operating system YC gives you, "Do these five things" or whatever, or is it more a philosophy of building companies? This is the confusing part for me as an outsider.

I think it's two major things. There is some of the operating system of what to do, but I think it was the philosophy of how to run companies: the idea of iterative deployment, technical people in charge, and being willing to bet on young people with a lot of energy and ambition but maybe less experience, throughout all levels of a company.

And then there was the related change to the whole ecosystem, compared to the pre-YC tech ecosystem. If we ran the clock back to 2004 and projected the technology forward to 2016, but not anything else about the shape of the startup ecosystem, what it meant to be an entrepreneur, how capital flowed, who got to run companies, all of those things, I do not think OpenAI would have been possible. The changes YC induced in the whole ecosystem: more leverage going to founders, young technical founders being able to raise lots of capital, the ability to work on ambitious things without a very proven résumé. Without those, I don't think OpenAI would have been possible. So this is a big change.

Is this all tied to the benefit you think there was in going from founder, to investor for a long period of time, back to founder?

There are all those benefits too. But I wouldn't say I really went from founder to investor to founder, because the first time I was a founder didn't really work out that well. Still, I started a company.

Learning from success

You learn some lessons from failure, but I think you learn way more from success.

No, hold on. We're not moving on from that. You've got to say more about it. There's some, I can't believe it, I think it's Anna Karenina, some great Russian novel, I'm very embarrassed not to know this, that starts with: all unhappy families are unhappy in their own way; all happy families are the same.

Yeah.

That explains why it jumped into memory. But I think this is really true. When I look at the lessons from where I've failed at something, I learn something generic about grit and determination, and something not to do. But most things don't work, and there are a lot of reasons why things don't work, so I think it's harder to put together the correct causation. And when I've had something really work, when I understand what parts of Y Combinator really worked or what parts of OpenAI really worked, trying to apply those lessons going forward has been much more helpful to me than trying to apply the anti-lessons of what didn't work.

So of course you should learn as much as you can from every data point: learn from the failures, learn from the successes. But in my own experience, when I've tried to apply those lessons, the lessons I learned from success were very good, and I should have applied them more. And the lessons I learned from failure were either fairly generic, things I already knew, or they got in the way of something else. I think this is generally true for a lot of people.

The power of repetition

Yeah, but isn't it that we already kind of know what we should do or what we should avoid, and it's the reminder, the constant reminder? The best description of my other podcast, Founders, that I ever heard is that it's church for entrepreneurs. If you really think about it... I used to drop in on Sundays, and I should go back to doing that. But really, I'm just telling the same... It's the same personality type appearing throughout history. Now this person happens to be building ships, and this person built technology, and they live in different times.

It's the same personality, that's for sure.

I'm kind of obsessed with the idea of things that last for a long period of time. The best companies can last a long time, but not as long as cities, and cities and countries don't last as long as religions. So out of all man-made things, what has lasted longest? I can't think of anything other than religion. There might be another answer. So I started studying. My mom was a fundamentalist Christian, so I was forced to go to church my entire childhood. And I started analyzing: what do all the main religions in the world have in common? We have a shared base of knowledge, usually some kind of book. We meet with like-minded fellow believers at regular intervals. And it's not like I go to church on Sunday and it's, "Okay, we talked about Jesus last week, but let's talk about this other guy." No: I go back to the same books and the same stories over and over again.

And I read your blog, and you even said something like this about the end of YC: you're repeating the same thing, you're telling it to them all the time, and then they leave the church, to use this analogy, and they stop doing it.

A hundred percent.

So it's not even the lessons. It's the constant reminder that this is important.

I extremely strongly agree with that. But I think it's better to be reminded of the positive things, like talk to your users more, ship products earlier, get more feedback, hold a higher bar for who you recruit and who you hire, and more quickly...

This is what I mean.

...but it's the positives that I think are good.

You have one of the greatest tweets. I have it saved on my phone. It's something like, "Skip the conferences, the fucking dinners, everything else." Essentially: make the product and sell the product. If you're not making it and you're not selling it... That's all you actually have to do. And again, it goes back to that simple genius.

So this is the other part I find most fascinating, because somebody asked me yesterday... Usually there's some historical equivalent for every founder I meet, where I can say, "Oh, that guy's kind of like Vanderbilt," or, "That guy's like Rockefeller, Ludwig," or any of these people. And they asked, "What's your historical equivalent for Sam?" I said, "There isn't one. I can't think of one." Because I don't know him well enough. I don't understand how he thinks yet.

What do you think now?

Well, this is the first of hopefully eight conversations, so I'll tell you in conversation seven.

All right.

But this is very rare. I just talked to Doug Leone, and he talked about one guy he hired, the founder of Nubank, who was an associate in VC and then left and founded one of the most successful companies. I said, "I've never heard of that. Doug, have you?" He goes, "No, that's the only one." So again, it's very rare. Mostly people go from founder, sell their business, unfortunately, and then become an investor, as opposed to running the business until they die, which is my preferred method of things.

What I'm curious about is this: you just said, "I learn more from successes." Well, the successes you saw came from being exposed to, what, 10,000 different companies in the decade or decade and a half you were doing this. And obviously you saw maybe the half-dozen or the dozen that were the best in the world. So are you taking their successes as instructive as well?

Yes, totally. And I think people do... You're an incredible student, but there are a lot of pretty good students of entrepreneurship, and people often try to look for those lessons, the things that really worked. And as you said, it's kind of the same thing over and over again, done in different industries, but you have to be reminded of it a lot, and it's unglamorous.

No customers, no product, no playbook

I had no understanding going into this conversation. I think I have a slightly better understanding now. I told you before we started that this is just for my own edification. But even with the people who influenced you, like Peter Thiel saying, "No, dummy." He's obviously going to say, "No, dummy, this is working. Why are you doing anything but the thing that's working?" So there must be examples where you're like, "Hey, I've given this advice to other founders a million times," and then you catch yourself: "Oh, shit. I'm not even applying my own advice right now."

Totally. I'll give many examples of that. But I think it's also instructive to ask: what was the new thing? What didn't you have the advice for?

The thing that was really different about OpenAI from anything I'd had pattern matching for before is that it was four and a half years from when we started the company to when we launched our first product.

The opposite of YC advice, right?

Yes. And although there were all these ways in which managing a research team was similar to selecting and advising founders, learning, to whatever degree we learned it, because I don't think we did it perfectly, how you manage through the part where you don't have the external signal from customers, and you're doing what would normally be the catastrophic startup advice of not shipping a product for four and a half years: that was very difficult. We tried all these things to replace the signal of "Do customers actually like the product?" with "Is our research actually working?"

One of the things that actually worked was during the Dota 2 days, when we were trying to use RL to beat people at a video game. We put up a leaderboard, and people could see how different ideas were performing. That was objective and real, and people wanted to go up it. But we had to try all these things to basically simulate end users, and that was a totally new and interesting problem I had no pattern matching for.

How did you work your way through that? What was your thinking?

We asked a bunch of people who had been at the great research labs of the past. OpenAI started at a time when everybody in Silicon Valley, as their vanity project, including me, wanted to start a research lab. There were all these very popular books about the heyday of Bell Labs or Xerox PARC. Everybody was talking about it; there was a huge amount of discussion. In fact, I see one of the books about Bell Labs over there. But there weren't a ton of people who had, in living memory, how to actually do it. So we talked a lot to Alan Kay. We talked to a handful of other people, and we got some advice from them about what made a really good research lab. Some of it was really good. Some of it didn't translate as well to the current moment.

Well, you also didn't have a giant monopolistic profit-printing machine. Bell Labs was spun out independently. Polaroid did a lot more research when it essentially had a monopoly on instant photography. I just read the biography of the founder of Honda. The guy created the most successful motor vehicle of all time, the Honda Cub, sold uninterrupted for 60 years, millions of fucking vehicles. And he arrived at the same conclusion Bell Labs did: he thought research and development had to actually be separate. It was spun out of the company and had separate ownership, just like Bell Labs.

We did not have that. In fact...

No, you did not.

The first day

When I think back to those early days, I mostly feel like I was trying and failing to raise money. That's my dominant memory of the early days of OpenAI. So much effort, so frustrating. I wish we'd had some sort of cash machine like that.

I remember one of my clearest memories of all of OpenAI. This is coming out of 2015: the first day was right after New Year's in 2016. Twelve of us, or eleven, showed up at Greg Brockman's apartment around 9:00, 9:30 on a Monday or Tuesday morning, something like that. Let's say it's January 4th. Everybody's there, it had been this big effort, and everybody walks in with a lot of excitement. It feels like the first day of school or whatever. And then very quickly people look around the room, and they're like, "Well, what do we do now?" Someone says, "Okay, we should get a whiteboard." Greg gets someone to go off and find a whiteboard. The whiteboard comes. Look around again: "What are we supposed to do now?" And you just feel the energy in the room collapse. None of us knew what to do.

It was not like building a product startup. It was not, "Let's build this product and talk to customers." It was, "Okay, we've said we want to make AGI. Maybe we should write some papers." "Okay, let's write some papers." "Maybe we should think about some ideas." "Okay, let's think about some ideas." Everybody has their moments of "I have no idea what I'm doing." That was one of mine. We had just launched this thing, and none of us had any idea what we were going to do.

So we did what we knew how to do, and eventually we figured out that a lot of things didn't work. Eventually we figured out a kind of rhythm for making and then evaluating research bets. Far from perfect, obviously, but we did find a gradient we could progress along, and we figured out how to get the resources that very smart people needed, and how to make sure we weren't completely getting lost in the wilderness. And over some number of years of mostly chaotic stumbling, we eventually made most of the big discoveries. What started as the unsupervised sentiment neuron paper turned into GPT-1, and then eventually GPT-whatever. The scaling laws work, which gave us the confidence not only to buy the compute but the understanding of how to scale up our models, came together, and many other things too. Through this process we learned things like the leaderboard idea, which worked. We also learned the incredible power of external demos for an eminent person the researchers really wanted to impress. And we learned a bunch of things that didn't work, like fake deadlines.

That has to be so disorienting to live through. You're twelve people in an apartment, you don't even have a whiteboard, you don't know what to do. Fast-forward a decade, and you have a billion people using...

A very strange experience.

Letters to his son

Do you keep a journal?

When my first kid was born, I would get home at the end of the day and rock him to sleep and just talk to him or whatever. I needed to come up with things to talk about, so I would tell him about my day, what we were struggling with, what I was worried about, what was happening. It was kind of fun for me to do, and I thought, "This is sort of interesting, and someday it'll be interesting for him to have." So I started writing to him. Every Sunday I would write him a letter. I would just talk, and then I would write it down. I only ever did eight of them or something.

How many kids do you have?

Two.

Bezos has this great line about building Amazon: "We're trying to do stuff that we can tell our grandkids about, that we're proud of." And those things are hard. The fact that you were writing to your son...

Oh, man.

Keep writing the letters. And if you don't, and this is real quick, just because I've read enough books about this: guess what? Founders don't write autobiographies when they're 40. They write them when they're 70, looking back, wishing they could do it again, and so much has been lost to the sands of time. They all repeat this: "I wish I'd journaled." So even if you don't do it, you have enough resources. What I would do is have a book written, even if it's for internal purposes only. Have you ever read The Little Kingdom by Michael Moritz?

I never read it.

Oh, you have to. It's the first six years of the history of Apple, written by Michael Moritz.

Isn't it crazy that he wrote that book?

He's a phenomenal writer.

He is a great writer. A beautiful writer.

A crazy writer who winds up being one of the best venture capitalists of all time. But the point is that the book ends before Steve has even been kicked out of Apple. So you get what actually happened. You're going to want this. You might not want it now, but you're damn sure going to want it when you're 60 or 70.

The thing that was so interesting was the mindset of writing to your kid. You really can't hide behind anything. I really care what my kid is going to think of me, so: this thing happened, I didn't feel great about it, I'd better do it differently next week. It was an extremely interesting mental framework. Maybe I'll find some way to do it again.

Not maybe. You're going to do it.

Okay.

Sam, thanks for taking the time. This was awesome, man.

Thank you.