Pod Insights Y Combinator Podcast Sam Altman: The Future of OpenAI, ChatGPT's Origins, and Building AI Hardware About this episode A fireside with Sam Altman at AI Startup School in San Francisco. Sam Altman grew up obsessed with technology, broke into the Stanford mainframe as a kid, and dropped out to start his first company before turning 20. In this conversation, he traces the path from early startup struggles to building OpenAI—sharing what he’s learned about ambition, the weight of responsibility, and how to keep building when the whole world is watching. He opens up about the hardest moments of his career, the limits of personal productivity, and why, in the end, it's all still about finding people you like working with and doing something that matters.
Transcript 0:00 Speaker 00 We said, okay, we're gonna go for AGI. Ninety-nine percent of the world thought we were crazy. One percent of the world that really resonated with, you know, in ten or twenty years, unless something goes hugely wrong, we'll have like unimaginable superintelligence. This is the best fucking time ever in the history of technology, ever, period, to start a company.
0:20 Speaker 01 Well, Sam, thank you so much for joining us and thanks for all the inspiration. I mean, OpenAI itself is a true inspiration for any really, really ambitious person. Maybe we just start with that. I mean, what were some of the decisions early that seemed small that turned out to be incredibly pivotal?
0:42 Speaker 00 I mean, just deciding to do it was a big one. Like, we got very close to not starting OpenAI.
0:50 Speaker 00 AGI sounded crazy. I was, I had Gary's job then and we were, you know, there was like all of this other great stuff to do that would work, all these great startups. And AGI was like kind of a pipe dream. And also, even if it was possible, DeepMind seemed like impossibly far ahead. And so we had this year over the course of 2015 where we were talking about starting it. And, you know, it was like kind of coin flippy. And I think this is the story of like many ambitious things where they seem so difficult and there's such good reasons not to do them that it really takes a core of people that like sit in a room, look each other in the eye and say, all right, let's do this.
1:28 Speaker 00 And those are like very important moments. And I think when in doubt, you should lean into them.
1:32 Speaker 01 So there were just a billion things, a billion reasons why people might say you shouldn't do it. I mean, off the bat, like even, you know, one of the things you figured out was the scaling laws.
1:42 Speaker 00 It's so hard to remember what it was like. Next year will be our 10 year anniversary. And so not, yeah, thank you.
1:54 Speaker 00 But to like, remember what the vibes were like about AI ten years ago, that was like way before the first language models that worked. We were trying to like play video games and we had this little robotic hand that could sort of barely do a Rubik's Cube. And we had no ideas for products, no revenue, no really idea that we were ever going to have revenue. And we were like sitting around at conference tables and whiteboards, trying to come up with ideas for papers to write. It was such, it's like hard to explain now, because it looks so obvious now, how improbable it seemed at the time. And how the idea of ChatGPT was like completely in the realm of science fiction.
2:34 Speaker 01 I mean, one of the things that really jumped out at me was you sort of, you know, rallied this idea that you should be working on AGI. And then simultaneously, you found the smartest people in the world who were working on that thing.
2:48 Speaker 00 That second part was sort of easier than it sounds. If you say we're going to like do this crazy thing and it's, and it's exciting and it's important if it works. And other people aren't doing it, you can actually like get a lot of people together. And so we were, we said, okay, we're going to go for AGI. Ninety-nine percent of the world thought we were crazy, one percent of the world that really resonated with. Turned out there were a lot of smart people in that one percent and you could get like, there wasn't really anywhere else for them to go. So we were able to really concentrate the talent and it was a mission that people cared about. So even though it seemed unlikely, if it worked, it seemed super valuable. And
3:27 Speaker 00 we've observed this many times with startups. If you are doing the same thing as everyone else, it is very hard to concentrate talent. And it's very hard to get people to like really believe in a mission. And if you're doing like a one of one thing, you have a really nice tailwind there.
3:42 Speaker 01 Okay, so some people in this room might be thinking like, should I try to start an OpenAI scale thing off the bat? You know, you also worked on Loop your first time around, you know, were there lessons from that?
3:55 Speaker 00 OpenAI was not an OpenAI scale thing off the bat. OpenAI was like eight people in a room and then it was 20 people in a room. And it was very unclear what to do and we were just like trying to write a good research paper.
4:08 Speaker 00 So the things that eventually become really big do not start off that way. I think it's important to like, dream that it could be big if it works. Nothing big starts that way and
4:21 Speaker 00 Vinod Khosla has this quote that I've always liked, which is, there's a very big difference between a zero million dollar startup and a zero billion dollar startup. But they both have zero dollars of revenue. They're both like a few people sitting in a room and you're both trying to, they're both just trying to get the first thing to work. So the only, the only advice I have about trying to start something big is pick, pick a market where it seems like there's some version of the future where it could be big if it works. But other than that, it's like one dumb foot in front of the other for a long time.
4:53 Speaker 01 How people use ChatGPT has changed a lot. How people use your API has changed a lot. What surprises you the most with the latest models like O three and what emergent behaviors or use cases are standing out to you right now?
5:06 Speaker 00 I think we're in a really interesting time. We haven't been in one of these for a while, but like right now, we're in an interesting time where the, the product overhang relative to what the models, like what the models are capable of is here, the products that people have figured out to build is way down here. There's a huge, even if the models got no better, which of course they will. There's a huge amount of new stuff to build. And also like last week, O three cost five times as much as it did this week. And that's going to keep going. I think people will be astonished at how much the price per performance falls. We have an open source model coming out soon. I think people are going to be, yeah.
5:48 Speaker 00 I don't want to like steal the team's glory and pre-announce this, but I think you all will be astonished. I think it will be like much better than you're hoping for. And the ability to like use it to run incredibly powerful models locally is gonna like really, really surprise people on what's possible.
6:05 Speaker 00 But so you have this world where like, the model capability has gone into like a kind of like a very new realm. The cost of the APIs are gonna keep falling quite dramatically. The open source models are gonna be super great. And I think we have not yet seen the level of new product innovation that the reasoning models are capable of, which makes sense. They're pretty new. But this is like an exceptional time to go build a company that takes advantage of this sort of new thing that exists. This sort of new square on a periodic table that no one has built with yet. So. Only in the last month, I think we really started to see startups that are saying, okay, like reasoning models are different. You know, the whole interaction model is different and really building for that.
6:48 Speaker 01 I mean, for me, even memory has turned into it. It feels like I'm talking to someone who knows me, which is interesting.
6:55 Speaker 00 Yeah, memory is my favorite feature that we've launched this year. I don't think most people at OpenAI would say that because we've launched a lot of stuff, but I love memory in ChatGPT. And I think it points to where we will hopefully go with the product, which is you will have this like entity that gets to know you, that connects to all your stuff and that is like proactively helping you. It'll, it won't just be like you send a message and it sends you one back, but it'll like be running all the time. It'll be looking at your stuff. It'll know when to send you a message. It'll know when to go do something on your behalf. You'll have, you know, special new devices and it'll be integrated on every other service you use and you'll just have this thing running with you throughout your life. I think memory is the first time where people can sort of see that coming.
7:37 Speaker 01 Uh, I, back in the day, you tweeted a little bit about her. What, you know, when is that coming? Can you give us an alpha leak around that?
7:45 Speaker 00 I think gradually is the answer. No, no, if I had a date in mind, I would probably just be excited and tell you.
7:52 Speaker 00 But like, it's a little bit here with memory.
7:55 Speaker 01 Right.
7:55 Speaker 00 It'll be a little more here when it's persistently running the background and sending you stuff. It'll be a lot more here when we ship the first new device. But I think the key of her is not the little piece of hardware. It's that this thing got to a point where it could run in the background and feel like a sort of AI companion.
8:12 Speaker 01 I guess we're starting to see the power of LLMs with integrations into your real data. You know, I've heard rumors that MCP is coming to OpenAI.
8:22 Speaker 00 I think today.
8:23 Speaker 01 Yeah. Oh, today?
8:24 Speaker 00 I think so.
8:25 Speaker 01 Fantastic.
8:29 Speaker 01 What has been surprising about the actual integrations? Like, have you been seeing people actually operating on their core database? You know, at YC, we actually have that agent infrastructure internally and we use it all the time.
8:41 Speaker 00 Definitely people are starting to use ChatGPT as this like operating system with everything, with their whole lives in it. And integrating into as many data sources as possible is important. Devices that are always with you, like new kinds of web browsers, the connection to all data sources, memory, and then a model that's persistently running. Put all that together, I I think you get to like a pretty powerful place.
9:04 Speaker 01 Do you think that'll be in the cloud in the future, or will it be on our desktop or some mix of both?
9:08 Speaker 00 Some mix of all of that. Definitely, people will run local models for some things like. Man, if we could push like half the chat GPT workload onto your local devices, no one would be happier than us. Like our cloud. We I think we will run the like largest and most expensive piece of infrastructure in the world pretty soon. So if we could push some of that off, that'd be great. Um, but a lot of it will run on the cloud.
9:30 Speaker 01 Is it surprising to you how hard it is to get compute?
9:33 Speaker 00 I mean, we've gotten really good at it. But it is.
9:38 Speaker 00 We we went from like a.
9:41 Speaker 00 No, a zero, like no chat GPT dot com didn't exist two and a half years ago to like. The fifth biggest website in the world, it'll be the third at some point, hopefully someday the first if our current growth rates continue. And I think doing that is just hard no matter what. Like that's, you usually get longer than we've gotten to to scale up a. Like infrastructure to for a new company, but.
10:05 Speaker 00 You know, there's like a lot of people that want to help.
10:08 Speaker 01 Well, it's incredible, incredible work that you guys have been doing. Um, we're seeing reasoning models like O three and O four mini. Uh, evolve in parallel with multimodal models like four O. What happens when these two threads converge and what's the vision for GPT five and beyond?
10:24 Speaker 00 I mean, we we we won't get all the way here with GPT five, but eventually we do want one integrated model that can like. Reason really hard when it needs to and generate like real time video when it needs to do that. If you ask a question, you could imagine it thinking super hard, doing some research, writing a bunch of code just in time. For like a brand new app only for you to use or kind of like rendering live video that you can interact with. So I think that will feel like. A real new kind of computer interface. The AI sort of somewhat already does, but when we get to a model that has like true, complete multimodality, like perfect video, perfect coding, everything and deep reasoning, that will, that will feel like quite powerful.
11:07 Speaker 01 It seems like that might be a hop step over to the embodied aspect. You know, that's having vision, having speech and having reasoning is a hop step to, you know,
11:19 Speaker 00 Yeah.
11:20 Speaker 01 basically the robot we want.
11:21 Speaker 00 Our, our, our strategy has been nail that first and then make sure we can connect that to a robot. But the time for the robot is coming soon. Um, I think I am very excited about a world where when you sign up for like the highest tier of the chat GPT subscription, we send you a free humanoid robot too.
11:39 Speaker 01 I mean, that future is going to be pretty wild, being able to have robots that do real work in the real world.
11:46 Speaker 00 I think we're not that far away now. Uh, the mechanical engineering of robots has been quite difficult. Uh, and, and the sort of AI for the cognitive part has been quite difficult too, but it feels within grasp. Um, and I think in a few years, robots will
12:04 Speaker 00 start to do super useful stuff. Making a billion robots is still gonna take a while, but I don't know. I'm interested in the question of how many robots do you need to fully automate the supply chain? Like if you make a million humanoid robots, the old fashioned way, can they run the entire supply chain? Drive the mining equipment, like drive the container ships, run the, you know, foundries and make the new robots. And then maybe like you actually can get a lot of robots in the world quickly, but the demand for humanoid robots in the world will be far more than we know how to think about with the current supply chain.
12:35 Speaker 01 I guess when you were sitting in my seat, one of the things you led was a, you know, a lot more investment into hard tech at Y C. Um, sitting here where we are geopolitically, you know, what, what do we need to do to make sure that America can actually have manufacturing and industrial capacity? You know, we can't even build precision screws and large sheet metal without crazy cost overruns, you know? What can we do to make sure that happens here?
13:00 Speaker 00 There are all of these answers that people throw around and have thrown around the same things for a while and it clearly hasn't worked. So, uh, I think all of the policy is worth trying, but my instinct is we need to try something new. We shouldn't keep trying the same failed stuff. And, you know, like AI and robotics does give us a new. Possibility of a way to bring manufacturing back here and to bring sort of these complex industries here in a really important new way. And I would say that's at least worth trying.
13:29 Speaker 01 Yeah.
13:31 Speaker 01 What does defensibility look like here? You know, one of the classic questions is, you know, how do I start a startup that doesn't get run over by OpenAI? That's sort of the number one question that's in our chat.
13:43 Speaker 00 We don't want to run you over.
13:48 Speaker 00 Look, we're gonna do our thing hopefully very well. We are gonna try to make the best super assistant out of ChatGPT that we can. We're gonna add the things that we think we need to add to that. But that is like one small part of the opportunity in front of us. And it makes us sad when people are like, I'm gonna start a new startup and I'm gonna like make a version of ChatGPT. Cause we think we're gonna do that pretty well. And we have like, you know, kind of a big head start. But
14:18 Speaker 00 there is so much more space to go after. And there are so many incredible other companies that have been built using our platform. We would like to make it easier for you all. We would like to do more things like finally now you can imagine that ChatGPT could drive a lot of traffic to new startups. And that there's like a new kind of app or agent or whatever you want to call it, store that we could do inside of ChatGPT and drive traffic to new startups to help. You could imagine that we could do like a sign in with OpenAI and people could bring their you know, personalized model and easily connect it to a new startup. And that would probably help in a bunch of ways. So we want to be a platform for other people to build stuff. Our advice is like, don't build our core
14:59 Speaker 00 you know, chat assistant.
15:02 Speaker 00 But there is another problem, which is and this is the same for every, every kind of like moment that I've seen in startup history. People get excited about the same thing at the same time. And so rather than go build the thing that you have thought of that is not everybody, what else is doing. We are like very social creatures and we get very influenced by what other people are doing. And I bet if Gary listed off the five ideas that he hears most often of what people want to build with AI, like half the room would raise their hands for working on one of those five. And And this, I, there is hopefully in this room, um, the person who's gonna start a company that is like much bigger than OpenAI someday.
15:45 Speaker 00 And I would bet that person is not working at any of the five. So it is hard to build something defensible if everybody else is trying to do the same thing. Sometimes it works. It's not impossible. But the best, the most enduring companies are usually not doing the same thing as everybody else. And that gives you time to figure out what the great product is, how to build the technology before you have to answer the defensibility question. It took us a long time to figure out how to answer the defensibility question for ChatGPT. Um, we had built this thing and for a long time, the only defensibility was like, we had the only product out in the market. Um, and then we kind of like be a brand that started to be well known. And now we have things like memory and connections, a whole bunch of other stuff that really is defensible.
16:29 Speaker 00 But, you know, that was like a fair criticism of us for a long time. We didn't have any defensibility strategy. We just like had the only good thing out there. And then you have some window before which you have to build defensibility.
16:43 Speaker 01 One of the things we've talked about in the past is that, uh, we both are big followers of, uh, Peter Thiel in that he talks a lot about being contrarian, but right.
16:55 Speaker 01 I think that you've.
16:57 Speaker 00 Peter is a genius.
16:57 Speaker 01 Absolutely. And you've found, you've been contrarian in really fundamental ways. I mean, going back, you know, to beginning of the conversation, people thought, oh, this idea that the scaling laws are valuable. Today it's, you know, taken as basic truth, but it was exactly the opposite of ground truth. Not that many years ago, when you got that pushback, you know, what did you and your team feel? Did you say, you know, F you, I won't do what you tell me, you know, I'm going, I'm going to push back against, you know, getting pushed back means that this is a contrarian area and we're gonna bet here and we're gonna be right.
17:33 Speaker 00 It is hard to have conviction in the face of a lot of other people telling you you're wrong. And I think people who don't who say it's easy are not being honest. It gets easier over time. But like I remember one time, I can say this one because it got publicized. In early, not early, a few years into OpenAI, where Elon sent us this really mean email. We'd been working on it for a while and said we had a zero percent chance of success. Like not zero point one zero that we were totally failing. We had showed him like GPT one recently. He was like, this is crap. It's not going to work. Doesn't make sense. And I was really a hero of mine at the time. And I remember going home that night and being like, what if he's right? Like this fucking sucks. You know, you're working so hard on this thing.
18:17 Speaker 00 Like you're pouring your life force into it. And you have these people who are smart and that you look up to. And they say you are totally wrong. Or, you know, this is just not never going to work. Or you don't have defense ability. Someone's going to kill you. This is going to happen. That's going to happen. And I don't have a magic answer other than it's really tough and it gets significantly easier over time. But it's going to happen to all of you. And you just like get knocked down and get back up and brush yourself off and try to keep going.
18:50 Speaker 01 Let's talk AI agents, you know, that's sort of the level three AGI. This is the year, I think Greg Brockman talked about recently. This is the year of the agent.
19:03 Speaker 01 With tools like operator code interpreter, what kind of workflows do you think will disappear? Or appear that we just aren't ready for yet?
19:12 Speaker 00 For a long time, ChatGPT was like a Google replacement. You could ask it something that was about as long as a Google query. You know, maybe like half an hour worth of Google queries it could assemble together. And that was still pretty good, but it didn't. It it it still felt like a more advanced version of search. But now you start to see things where you can like really give a task to Codex, for example, or to deep research. And you have this thing go off and do a bunch of stuff and come back to you with like a proposal. It's it's like a very junior employee that can work on something for like a short period of time. And if you think about how much. Of the work that the world does is work that can be done in front of a computer
19:54 Speaker 00 in like few hour chunks where you then have someone say like, okay, that was good enough or not. It's quite a lot. So I, I think this is gonna go, this is part of that overhang we were talking about earlier. But I think this is gonna go quite far. And I think with current O three, to say nothing of our next model, you can build a lot of experiences like this.
20:12 Speaker 01 How do you see the future of human computer interaction and interfaces? And what are sort of the limitations of those interfaces that, you know, motivated you?
20:22 Speaker 00 You know, one of the things that I think sci fi got right is
20:27 Speaker 00 the idea of the interface almost melts away. Like, voice interfaces today, we think of as something that is kind of sucky because they don't work that well. But in theory, if you could say to a computer, this is exactly what I want to happen today. And if there's any changes, if like I'm delayed, if you know, something happens, I trust you to like go off and do all those things, but like, I don't want to be interrupted. I don't want to think about it. Um, and it just did it all and you trusted that it worked. That would be an interface that almost melted away, except when it, you know, was like a super great human assistant needed to talk to you. But you would be like really thrilled. When I like use my phone today, I feel like I am like walking down Times Square in New York, getting like bumped into by people. I love my phone. It's an incredible piece of technology.
21:11 Speaker 00 But it's like notification here, this thing happening, you know, this thing popping up like bright colors, like all kinds of flashing things in me.
21:18 Speaker 00 It's just stressful. And I can imagine an interface where the computer mostly melts away. It does the stuff I need. Um, but I really trust that it's gonna do a great job of like surfacing information to me, making judgment calls about when I have to do, acting on my behalf when it should. Uh, and I'm, I'm quite excited for that. Uh, I'm not gonna like tell you what the new device is. Well, I'll tell you like one on one, but I'm not gonna tell. But it's very like, I, I hope.
21:46 Speaker 00 I hope we can sort of like show people a different way to have computers.
21:51 Speaker 01 Is that one of the reasons why you brought on one of the greatest living designers on the planet in Johnny Ivan I O?
21:57 Speaker 00 Yeah, he is. He is amazing. He is. He really lives up to all the hype. I think we've only had kind of two big revolutions in computer interfaces. Really, in the last like fifty years, we had the, you know, keyboard and mouse and screen, and then we had touch and phones. And the opportunity to do a new one doesn't come along that often. And I think AI really does totally open the playing field for something completely new. And I think if you've got to pick one person to bet on to figure that out, he's the obvious bet.
22:28 Speaker 01 So one of the things that we've been debating at Y C that, you know, don't know if this is good. Might be scary for a lot of software engineers who want to create B two B SaaS is this idea that what if in the future you had your, you know, underlying database, you have an API layer that is, you know, your access control and enforces your business logic. And then the interface is the L L M, like your computer is literally, you know, the agent. And you have just in time software, there are like complex flows. You're just gonna go in and it'll code gen an artifact or, you know, a pain for you that like does that thing you wanted. And it'll go in the file and it'll bring it back if you ever need it.
23:10 Speaker 00 That's gonna happen.
23:11 Speaker 01 Yeah.
23:11 Speaker 00 Look, here, here, here, there are two ways you can look at this. First of all, assume you all are like starting startups or have started startups, think about starting startups. This is the best fucking time ever in the history of technology, ever, period, to start a company. Yeah, this is.
23:31 Speaker 00 And but part of the reason it's the best is because like the ground is shaking and it's true. There are a lot of these challenges. So on one hand, you can look at something like that and say, we have been a, you know, SaaS company. And now like all of the code can just be generated right in time when someone needs it. And what does that mean for us? Or you can look at it and say, wow, this is gonna happen, but it's gonna happen to everybody. And the way startups win is.
23:59 Speaker 00 When they can iterate faster than big companies and they can do it at a much lower cost, like big companies have a lot of advantages, but they iterate very slowly. And they, you know, if something is like very cheap, then a lot of their big advantages go away. So you can look at this, all of these problems, one way or another, but the way I would recommend looking at them is everybody is going to face the same challenges and opportunities. But when the clock cycle of the industry changes this much, startups almost always win. And we've probably never seen it change this much. Act on it from that direction. I think you'll be in incredible shape. Maybe you can invite me sometime to do a talk about like what the areas of defensibility
24:44 Speaker 00 that you can build over time are, because I think that is the inherent question. People are like, oh, okay, you know, I'm a SaaS company. There's going to be just in time software. I think the question behind the question is like, what are actual defensibility strategies? So that would be a fun talk someday.
24:57 Speaker 01 I guess, you know, backstage at one of the last events we had, you know, we were talking about the, there's this book that's sort of like the classic McKinseyism, which is the seven powers. And I was just thinking about that, like, I never would have thought like the two of us technologists sitting around actually citing a book that, you know, McKinsey consultants are known for.
25:17 Speaker 00 Feels so wrong.
25:18 Speaker 01 Yeah, I don't know.
25:19 Speaker 00 Aesthetically feels terrible, but yes.
25:21 Speaker 01 Seven powers, I guess. We're entering this age of intelligence. I love that essay of yours. What do you think this era will mean for, you know, how we live, how we work, and how do we create value for each other as a society?
25:37 Speaker 00 You know, in some sense, the whole arc of technology is one story, which is we discover more science, build better tools, all of society like builds the scaffolding a little bit higher. And we, we have this more impressive tool chain and the whole point of it is that one person can do way more than they could before. And this has been going on for a long, long time, each generation. Certainly, I mean, if you compare a person today from a person a hundred or thousand years ago, one person is incredibly more capable. And the kind of like social contract is that you put something, you know, you build the next layer of scaffolding.
26:18 Speaker 00 But what someone can do now with this new set of tools, with this, you know, this new layer that's been built in is pretty incredible. And I think the one of the things that will feel most different about these next ten years versus these last ten years is how much a single person or a small group of people with a lot of agency can get done. And that is a bigger deal than it sounds like, because coordination costs are huge. And when we can empower people with more knowledge, more tools, more resources, whatever. I think we won't just see like a little bit more stuff get built. But because of these kind of coordination costs across people, we'll see like a real stuff change.
27:03 Speaker 00 So I think the the amount that one person or a small team get done, the satisfaction in doing that. And most importantly, like the quality of stuff we'll all get for each other will be quite remarkable. When I think back about the OpenAI story, I often think about just the kind of
27:20 Speaker 00 key few tens of people that did the amazing work that led to what we all have now.
27:26 Speaker 00 But I try to remember that I always also have to think about like the tens of millions of people, maybe it's more throughout history that started like digging rocks out of the ground, figuring out how semiconductors work, building computers, building the internet and on and on and on. That let this small group be able to work at such a high level of impact that they never would have been able to do without the collective output of society.
27:46 Speaker 01 Is it surprising to you to what degree, I mean, this room is, you're preaching to the room of the converted. But uh
27:53 Speaker 00 This is awesome, by the way.
27:54 Speaker 01 I mean, this is like the collected set of people who are gonna go create the future. But um, there's, you know,
28:00 Speaker 00 Yeah, there's probably, there's maybe like never been a gathering like this in one place before. This is, this is very cool to see.
28:06 Speaker 01 But at the same time, you know, we're in some ways, uh, this is the leading, cutting edge of all of society. Because there are seven and a half billion people who probably, you know, have not even tried this stuff yet. And not only that, their main interaction with it is that it doesn't work, that it hallucinates. You know, what do you have to say to the 3,000 people in front of you right now? Just, this is the thin edge of the spear. We are literally teaching people and giving people this technology.
28:39 Speaker 00 First of all, that's like a great place to be in. One of the most fun things about working at YC is you get to like live on the leading edge. And you get to be around the people who are the advance guard. And that's just like a fun way to live your life and you get to see what's coming and you know, hopefully have some small amount of input into shaping it.
28:59 Speaker 00 But I don't know, I think AI is like somewhat mainstream right now. The negative, the way that it's not is most people still think of AI as ChatGPT. And a lot of people use ChatGPT, but they use it like a chatbot. And they have not yet wrapped their head around what's coming next and probably you all have. But I don't know, it's like a great privilege to get to live a little bit in the future and you know, go build stuff for everybody else coming along.
29:27 Speaker 01 So you're sort of one of the best people in the world at bringing together the smartest people. What are some of the hardest lessons you've had to learn about hiring? A lot of the people in this room, like they have never managed a person before, let alone gotten someone to quit their, you know, six to seven figure job at some big company to come work on their revolution.
29:51 Speaker 00 Hiring really smart people who are clearly really driven and really productive and can work as part of the team, I think does get you
30:06 Speaker 00 90% of the way there. And the degree to which people focus on other things to hire for always surprises me. So I think, you know, given that we can't do the full 45 minutes right now, really smart people. Driven, curious, self motivated, hardworking, like good track record of accomplishment and can work really well as part of a team and sort of aligned with the company's vision. So that everybody's at least going for the same, the same direction. That works pretty well.
30:40 Speaker 01 I mean, by a strong track record, do you mean the person who's like, you know, sort of been an administrator and had like the, the, you know, top name at the top institution for 20 years? Or do you mean like, cause you, you went the other way.
30:55 Speaker 00 I, I don't, especially early in a startup, I don't believe in hiring those people. Their experience is valuable and there are times where you, you really need that. But I have not had success and to be frank, like Y C has not had that much success. Trying to start with like the very senior eminent administrator as one of the like, you know, as the first hire in a startup. I would, I would take like young, scrappy, but clearly like get stuff done. Over the person who has like the extremely polished track record. There will come a time where you need some of those people later, but I don't know how you do it.
31:37 Speaker 00 But when I, when I was like reading Y C applications, I would like never look at the resume items. You know, you worked at like Google or went to this college and I never cared. I would always go right to like, what's the most impressive stuff you've done? And then sometimes I would like not be convinced by that and go look at the resume. But that was always like a backup to me as a secondary thing.
31:55 Speaker 01 So they sort of look, look at what they've actually, what they've coded, what they've built, like their velocity, how they think about problems and solve them.
32:04 Speaker 00 I see P B back there. He has this quote, I hope it's his quote, because I've attributed him a bunch of times of hire for slope, not why intercept. And I think that's just like unbelievably great advice.
32:14 Speaker 01 Talk about being CEO of OpenAI. What are some of the hardest lessons there just overall?
32:20 Speaker 00 I don't recommend it.
32:23 Speaker 00 No one single challenge would be that hard, but the number of things we have to do at the same time and the kind of like number of other big companies that are gunning for us in various ways. It's just like more context than I thought it was possible to handle at once. And more sort of like switching from like big, big decision to like totally unrelated, but also huge decision.
32:44 Speaker 01 Looking ahead ten to twenty years, what are you sort of most personally excited about, you know? And what should people be building now to make that future possible? You know, there are people who are scientists, there are people who are software engineers, there are people who are, I mean, this is an all-technical crowd.
33:06 Speaker 00 Look, there's a lot. You know, ten to twenty years, unless something goes hugely wrong, we'll have like unimaginable superintelligence. And I'm very excited to see how that goes. Forced to pick one thing to just not leave it as like a vague answer.
33:22 Speaker 00 I think AI for science is what I'm personally most excited about. I am a believer that to a first-order approximation, all long-term, sustainable economic growth in the world, like everything that leads to people's lives getting better, is basically discovering new science and having reasonably good governance and institutions so that that science can get developed and shared with the world. But if we could vastly increase the rate of new scientific discovery with AI, I believe that would compound to just incredible increases and wonders for everyone's lives. So I think I'd pick that on that timeframe.
33:58 Speaker 01 I guess one of the things I've been always really impressed by is, you know, you personally recruited Helion to come do Y Combinator. And they're doing incredible things over on the fusion side. Was that something that you were thinking about even all the way back then? Or, you know, obviously, energy and climate was sort of a part of, you know, what everyone's worried about even back then, but...
34:24 Speaker 00 This is a little bit embarrassing. I've been obsessed with energy and AI as like the two, the things that I thought would be the two most important things, or at least the ones I was gonna be most that I felt most passionate about for a long time. And really like the two areas that I knew I wanted to like concentrate time and capital towards. I cannot recall ever thinking until like after starting OpenAI that they were gonna be so obviously related. That, you know, the that energy would be the eventually the fundamental limiter on how much intelligence we could have. And I don't know if I missed that because I usually am good at thinking about things like that. But I really did think of them as like very independent. You know, we were going to need AI to have all the ideas, energy to make all this stuff happen in the world.
35:08 Speaker 00 And I obviously, right after starting OpenAI, I got obsessed with meaning energy for AI. But like pre 2015, I think I thought of them as orthogonal vectors.
35:17 Speaker 01 I mean, I'm sure you've seen that chart that, you know, all the effective accelerationists in the room have seen around basically having a high standard of living, like the sort of really.
35:29 Speaker 00 Yeah, I'm obsessed with this chart. I've been obsessed with that chart a long time.
35:33 Speaker 01 It's directly related to the amount of energy that any given person has access to.
35:37 Speaker 00 Yeah, I I think this is one of the most amazing charts over a long, like, long, long period of human history, is the correlation of quality of life and abundance of energy and cost of energy. So that was, that chart and charts like that were a significant reason that I got obsessed with energy in the first place. It is, it is just this like crazy high impact thing.
35:59 Speaker 01 It sounds like it wasn't entirely interdependent. It was more you had twin interests.
36:04 Speaker 00 Yeah.
36:04 Speaker 01 You've literally woven them together.
36:06 Speaker 00 I had like the one interest of like radical abundance and just like what, what, what were the kind of technological leverage points to just like make the future wildly different and better? And these is the two kind of key things for that, but not as, not as much as the same vector. Now I think a lot about like, how much energy can we actually build on Earth before we just heat the planet too much from running the GPUs? And like, how long can we go before we have to put all the GPUs in space? But at the time, yeah, I really thought of them differently.
36:34 Speaker 01 I mean, it seems like one of the defining beliefs that technologists uniquely ideally have, that they believe that we can actually create that sort of abundance. You know, if you have intelligence on tap and then you have energy on tap, then how does that go? It's like, you know, all, all watched over by machines of loving grace.
36:56 Speaker 00 I I've never been to one of those degrother conferences in Europe or whatever. But I've always kind of wanted to go to one.
37:03 Speaker 01 This is the anti-degrowth.
37:04 Speaker 00 This is the anti-degrowth conference, totally. But I would like love to be like sitting, you know, in the dark and the cold with no one pulling out their phones and just like talking about how horrible everything was and there was no hope like I would love to experience that mindset once because I've never felt it. Um, and I think it is like it it is one of the movements that has been ever hardest for me to identify with. Obviously, this is like my crew and my world, but the the sort of like the optimism of startups of San Francisco, of the technology industry, of AI, of what all y'all y'all will do. Like, that is, that is like the natural space my brain abides, and it's very hard for me to
37:46 Speaker 00 really empathize with the other side of that, but I'm pretty sure we're right and they're wrong.
37:50 Speaker 01 How do we get there, though, right? You're this incredible vision of technology actually creating for abundance for others. I mean, you've already done so much, but you know, point us the way like you're how how else do we get there? How do we make it faster? You know, does government play a role in this?
38:09 Speaker 00 Almost just about five years ago, like pretty much this week, we put GPT three into an API. And people started playing with it. And it was barely usable. It was quite embarrassing. And in five years, we have gone from this like thing that could barely write a sentence to a thing that is like, you know, PhD level intelligence in most areas.
38:33 Speaker 00 Five more years, I think we'll be able to maintain the same rate of progress. And I think if we do that, if we also build out the infrastructure to serve that to people, then everybody in this room will figure out how to take that technology and adapt it to what everybody needs. The the analogy I like most for AI is the transistor, like the historical tech technical analogy. You know, some people figured out like a new really important scientific discovery.
39:02 Speaker 00 And society, the economy, whatever you want to call it, just got to work. Just did its thing. The magic of that just figured out how to make incredible value for people. And really, over a fairly short period of decades, significantly ramp up quality of life. I think this will be even faster and steeper than that, but I think it'll go in direction the same way. You know, we need to make the great technology, figure out the remaining scientific stuff, which I don't think there's much left. We need to figure out how to build out the infrastructure. That you all will need to be able to service. And then you all have got to go figure out what what people in the world need with this new magic.
39:38 Speaker 01 So let's flashback to two thousand and five, the very first batch of Y Combinator. How did you hear about Paul Graham? You were reading his essays.
39:47 Speaker 00 I was reading his essays. So I'd heard about like, he kind of had this cult following on the internet. But I heard about.
39:54 Speaker 00 What was then called the Summer Founders Program, and now it's just called Y Combinator. From Blake Ross, who I live in the same freshman dorm with and posted about it on Facebook.
40:05 Speaker 01 And then I think Paul said, oh, you're a freshman, you know, that there's like another batch coming. And what did you reply to him by email with?
40:14 Speaker 00 You know, funny you bring that up. I just dug up the email like a couple of days ago because I I felt I had been misquoted over time.
40:21 Speaker 01 I'm curious.
40:22 Speaker 00 And. The the his telling of the story is I said, like, I'm a sophomore and I'm coming. But I wrote a much nicer thing. It was like, oh, maybe there was some misunderstanding, you know, actually, I'm a sophomore and I can still make it and I would like love to if that's still okay to come the next day.
40:38 Speaker 01 So I, in some ways, you know, the wild thing is you're sitting in front of three thousand people who. Kind of was that, you know, they they are sitting where you were back in two thousand and five. What would you say to, you know, the Sam Altman from that time? You know, given what you know, all the things you've seen, all the things you've learned since, like, what are the things that you're most surprised you didn't know that? I mean, it just took, I mean. You've been through it, you know, like you've you've done it.
41:09 Speaker 00 I wish someone had like taught me the importance of like conviction and. Resilience over a long period of time. People don't really talk about how hard that is. It's it's like easy for a little while, but your reserves kind of like wear down on it. And how to how to keep that going for a long period of time.
41:30 Speaker 00 Also, just sort of like trust that it's eventually gonna work out. Like, obviously, my first startup didn't work that well.
41:38 Speaker 00 I think a lot of people kind of give up after one failed startup. But startups don't work out all the time. And learning how to keep going through that, keep working through that is, is, I think, really important. Developing like trust in your own instincts and increasing that trust as you refine your decision making and instincts over time. I think that's really important. Kind of courage to work on stuff that is out of fashion, but is what you believe and what you care about. I think that's really important. I had a kid recently, and the thing everyone tells you when you have a kid is that it is the best thing you will ever do. But also, that is the hardest thing you will ever do. Like, the, the good parts are much better than you can imagine. The hard parts are much harder. That is all totally true.
42:22 Speaker 00 And that is also basically what I feel like being an entrepreneur is like. The good parts are really great, better than you think. And the hard parts are like shockingly much harder than anyone can, can express in a way that makes any sense to you. And you have to just keep going.
42:36 Speaker 01 Sam Altman, everyone.
42:38 Speaker 00 Thank you.
42:41 Speaker 00 Thank you.
Previous Elon Musk: Digital Superintelligence, Multiplanetary Life, How to Be Useful Next Satya Nadella: Microsoft's AI Bets, Hyperscaling, Quantum Computing Breakthroughs Page updated October 12, 2026 · transcript and summary are machine-generated
