Transcription
So far, we've got a consumer business, a B2B business. There's this whole Johnny IV thing, which I'm sure, you know, we can't really talk about.
Johnny, Johnny, we got to start. I got to start over. I can't do that. So, leave that in, please. No, no, no. We're going to cut. All right.
Today, I'm here with Sam. Sam, before we start, do you have anything you need to say? You're my literal podcast bro now. Wow. This is great. How did it even come to this? It's so sad. You start a company, then you start being a VC, and now I'm here. Are you disappointed?
Well, I went the other way.
What do you mean?
Well, I was like a VC, then I did a podcast and now…
Yeah, it's been good for you.
It's great. I'm really proud of you. Okay. So, but I think this is great for you.
Thank you.
Okay. I think you're an incredible podcast. It's a very nice sweater, too.
Thank you. Thank you. Okay. So, uh… I want to start by talking about the… Stop.
What are you going to say?
Go ahead. I'll say it later when we're done recording.
I wanted to start by talking about the future of AI, and I'm… I want to talk about the medium-term because the short term is not as interesting to me. The long term, who knows? But like 5-10 years out is what I'm most interested in talking about. And I kind of want to try to pull out from you your best guess of a bunch of specific things. One of the places I wanted to start was in software. It seems like the most effective use cases so far, which I'm curious if you agree with, but seem to be um, coding and then chat and code.
Yeah, chat and code. I'm curious what's next, like on the next sort of… what's the next set of things right after that that will come?
Well, I think there will be incredible like other products; like there will be crazy new social experiences. There will be like Google Docs style AI workflows that are just way more productive. You'll start to see like you'll have these like virtual employees. But the thing that I think will be the most impactful on that 5- to 10-year time frame is AI will actually discover new science. And this is a crazy claim to make, but I think it is true. And if it is correct, then over time I think that will dwarf everything else.
Why do you think it'll discover new science?
Well, I think we've cracked reasoning in the models. We have a long way to go. I think we know what to do. And you know, 03 is already like pretty smart. You hear people say like, "Wow, this is like a good PhD."
What does it mean to crack reasoning?
The models can now do the kind of reasoning in a particular domain you'd expect a PhD in that field to be able to do. In some sense, we're like, "Oh, okay. the AIs are like a top competitive programmer in the world now," or "AIs can get like a top score on the world's hardest math competitions," or "AIs can like, you know, do problems that I'd expect an expert PhD in my PhD in my field to do," and we're like not that impressed. It's crazy, but it is sort of a crazy thing, you know. This reasoning ability of the models over the last year…
Are you surprised?
Yes, yeah. You thought that it was just going to be like the next token type. I thought it was going to take a little bit longer to get where we are now. The last… the last year of progress has been faster.
Did the way reasoning happened happen the way you thought it would happen?
Like often has happened in the history of OpenAI, sometimes pretty often, the dumbest first approach turns out to work. So I don't think I should be surprised by that anymore. And yet it's like a little surprising each time.
So reasoning will lead to science going faster or just new stuff or both?
Both. I… I mean, you already hear scientists who say they're faster with AI. Like we don't have AI maybe autonomously doing science, but if a human scientist is three times as productive using 03, that's still a pretty big deal.
Yeah. And then as that keeps going and the AI can like autonomously do some science, figure out novel physics…
Is it all that happening as a co-pilot right now?
Yeah, there's… there's definitely not like… you definitely can't go say like, "Hey, ChatGPT, figure out new physics," and expect that to work. So it is… I think it is currently co-pilot like… but I've heard like anecdotal reports from biologists where it's like, wow, it really did figure out an idea. I had to develop it a little bit more, but it made like a fundamental leap.
Yeah. Will it be easier to get AI to build you a whole business, like build you a whole e-commerce business, or like do like a hard piece of science or medium? I wonder about this. If you could like build AI a hundred-billion-dollar particle accelerator…
Yeah.
…and say you make the decisions. You look at the data, you tell us like, uh, you know, what experiments to run and we'll go like find the stuff and do it. So you spend a hundred billion dollars doing that or you spend $100 billion like building it infrastructure to like connect into the economy. Which of those will it have an easier time doing something remarkable with?
Yeah. And I think the physics is a cleaner problem.
Mhm.
You know, I think if you could get like new high-energy physics data and then AI, the ability to like run experiments, I think that's like a… a cleaner problem. Um, I've heard people say that they expect the first area of science, this is like… I don't know if this is accurate or not, but I've heard people say that they expect the first area of science where AI makes autonomous new discoveries to be astrophysics because there's just mountains of data and we don't have enough PhDs to look at it. And maybe it's not that hard to figure out new stuff, but I don't really know.
Okay, so science gets better; the coding and chat will just keep getting better. Does that lead on the business side? Do you also then have the ability to just… can you prompt a whole business? Like could you just say, "build me this type of business," it looks like this will happen, that will happen? There's already people doing that for like small stuff, you know. You hear these stories of people who like use AI to do market research and like figure out new products and then like email some manufacturer and get some dumb thing made and sell it on Amazon and run ads. Like there are people that have actually figured out at small scale, in the most boring ways possible, how to like put a dollar into AI and get the AI to like run a…
Mhm.
…toy business, but it's actually working.
Yeah. So, that'll climb the gradient.
Yeah. What about in the world of like physical stuff? 'Cause like I get that… I mean, it seems to me, you know, very clear that like software is just going this direction. Science, I know less. Take your word on it. What about like moving physical things around?
Behind, but I think we'll get there. Uh, for example, I think we have some new technology that could just do self-driving for standard cars way better than any current approach has worked. And that might not be quite what you meant by like humanoid robots, but if our AI techniques can like really go drive a car, that's still pretty cool.
Yeah. Um, humanoid robots are the dream obviously. I really care about that. I think we will get there eventually. It's been like a hard mechanical engineering challenge. That's more the issue.
No, the both things are hard, but like even if we had the perfect brain right now, I don't think we have the body yet. Um, we… we actually very early on at OpenAI, we used to work on this robotic hand, and it was hard for all the wrong reasons. Like the thing just broke all the time. The simulator was like a little bit off.
Wow.
But you know, we'll get there.
Yeah. I think 5 to 10 years we'll have great humanoid robots.
Yeah.
Like amazing. And they'll just like walk down the street, be doing stuff.
Yeah. I mean, you would think that's where a huge amount of step change unlocks, right?
I think that will be one of the moments that not only is unlocks a bunch of stuff in the world, I think that will feel the strangest. We get used to a lot of things. We get used to like ChatGPT doing these things that would have sounded like a miracle 5 years ago. But if you walk down the street and it's like half robots, are you going to get used to that one right away? I don't know. Probably you do. But it feels like a big difference. That's the one that'll feel like there's like a new species taking over us.
Yeah. I think that'll feel… I don't think it'll feel like a new species or that it's taken over, but I think it will feel like the future in a way that ChatGPT still does not. I think also if we can figure out great new computing devices to make that will feel maybe like the future, but as amazing as ChatGPT is or these new coding agents and they are amazing, it's like still stuck in the form factor of the past.
Yeah.
It's also stuck in… it's stuck in the computer.
Yeah.
Yeah. There's definitely something about that. It only can do stuff at a computer, but I don't know. Like, how much of the economic value in all the world do you think is like cognitive labor that can be done behind a computer?
Like half.
I was going to say a quarter, but maybe half. I don't know, but some big number.
Yeah.
Does stuff get much riskier once we have like super embodied intelligence? 'Cause those things are going to be way stronger than us, too.
I don't know about way riskier. I think like the ability to make a bioweapon or like take down a country's whole grid, you can do… you can do quite damaging things without physical stuff. It gets riskier in like sillier ways. Like I would be afraid to have a humanoid robot walking around my house that might fall on my baby unless I like really, really trusted it.
Yeah. What… what like… um, if you're thinking about… you know, we're back here in 10 years having another conversation and we're like, "Did AI do what we thought it would do?" What metrics are you expecting? Like is it that like the GDP growth curve has a kink in it? Is it that like life expectancy is up? Is it like there's less poverty? Is it something completely different?
So every year before the last, like maybe up until last year, I would have said like, "Hey, I think this is going to go really far, but it still seems like there's a lot that we've got to figure out." I feel very confident at this point, the most confident I've ever felt that we kind of like know what to do to get to incredible AI systems that are just super, super capable. If something goes wrong, I would say like somehow it's that we build legitimate super intelligence and it doesn't make the world much better. Doesn't change things as much as it sounds like it should.
Seems like a crazy thing to say.
Yeah.
Yeah, but like… I don't know. If I told you in 2020, maybe I did tell you, like, "We're going to make something like ChatGPT, and it's going to be as smart as a PhD student in most areas, and we're going to deploy it, and you know, a significant fraction of the world is going to use it and kind of use it a lot." Maybe you would have believed that, maybe you wouldn't have, but conditioned on that, I bet you would say, "Okay, if that happens, the world looks way more different than it does right now."
Yeah.
So it's like we have this crazy thing…
Yeah.
I mean, there's this thing where it's like the Turing test, you know, everybody's thought it just went by…
Yeah.
Nobody really cared, and…
Yeah.
I don't know what explains that.
Well, the fact that you can like have this thing do these… this amazing… amazing stuff for you and you kind of live your life the same way you did 2 years ago…
Yeah.
…kind of work the same way you did 2 years ago…
Do you think that's possible that we have like crazy super intelligence that's like 400 IQ and we're still… that's still the case?
I totally think that's possible.
Wow.
If it's off discovering new science for us, eventually society will figure out how to deal with that, but it may be very slow.
Well, what's funny is if it looked like a co-pilot, you'd kind of still credit the PI at the lab who was using this, you know, 400 IQ agent behind it. I think you kind of will no matter what. Humans are so wired to care about other humans. We need people in the story, you know? We need to talk about like, "That guy did that thing," or like, "made this decision," or "made this mistake," or had, you know, this whatever. That's why I was surprised that you don't think if we had like a super accurately embodied robot that we wouldn't start to imprint some of that on that robot.
I think we are. We will find out. I may be wrong. I think we will have more of a relationship than we do now as the thing is more embodied. But I think we are so deeply hardwired to care about other people, and that is going to turn out to be pretty deep in biology, and if you know it's a robot, no matter how humanlike it seems in other ways, you're not going to care that much. That's… that's speculation.
So reasoning was like one of the components of intelligence that like sort of got figured out. Is there like another thread going around like a topic of like agency or like some concept of like self-directedness? Like is that a thing?
The ability to work on a goal over a very long time with a lot of complicated steps along the way, I think is maybe what you're going for.
Yeah, same thing. And that is definitely something we're working on.
Yeah. What about the future sort of technical path? Would you say is now inevitable and what parts would you say you're still not sure which way it'll break?
I think we will get to extremely smart and capable models, uh, capable of discovering important new ideas, capable of automating huge amounts of work. But then I feel totally confused about what society looks like if that happens. So I'm like most interested in the capabilities questions, but I feel like maybe at this point more people should be talking about like, "How do we make sure society gets the value out of this?" I think those questions have somehow become harder and less clear. This… I mean, this is a crazy statement that like we're going to solve the super intelligence, but maybe society still sucks.
Yeah.
Yeah, I can't… feel right to me, but like I can't tell if… sometimes with some of these statements that people don't react because they only kind of believe it, and maybe that's part of why, but I agree. I mean, that's what, you know, the history of a lot of this stuff has been like… get said… people don't quite believe it, and then it happens, and then people just kind of adjust. So, I don't know what to make of all that either. I feel like… I feel like we've been very right on the technical predictions, and then I somehow thought society would feel more different if we actually delivered on them than it does so far. But I don't even… it's not even obvious that that's a bad thing.
Well, one of the more obvious short-term impacts will be potentially… or you'd think would be like employment. That seems like something that like we don't need like… you know, we don't need to believe crazy leaps to see that there ought to be some impact. Like this is going to happen in customer support very obviously right now, for example.
Yeah. I mean, my take on this is a lot of jobs will go away; a lot of jobs will just change dramatically. But we have always been really good at figuring out new things to do and ways to occupy ourselves and status games or ways to be useful to each other. And I'm like not a believer that that ever runs out.
Yeah.
Uh, I think now I do think it gets maybe sillier and sillier looking from our current perspective.
Yeah.
Like podcast bro was not a real job not that long ago, and you figure out how to monetize it, and you're doing great, and we're all happy for you.
We're so happy.
But would… would like the subsistence farmer look at this and say this was a job or this is you like playing a game to entertain yourself?
I think they would subscribe to this podcast. I bet they would. They would like it. But I do think that there is like a big issue here in the short term. I think long term, who knows? I mean, one of the things I'm curious about is in general, so like going from a time when like everybody was a farmer and like nothing that we're currently doing makes any sense to… like now you have all this stuff. Is it different this time if there's enough resources to go around? Like at some point, are there enough resources to go around and like that's what creates the difference where now people just like don't make new jobs? And this… so Venode was on the podcast last week. He… we're going to release this one first, so this won't be out by the time that this is said, but his point was that people will just consume a lot more leisure. So he kind of took the view of like, "I think this time like there's just going to be an abundance of resources. Everybody's going to have the stuff that they need; we're going to be able to build, you know, buildings, and people can now just like enjoy their lives." Again, I think like the relativistic framing matters here to us. I bet it will look like those people are just consuming crazy amounts of leisure. I guess this looks like a lot of leisure, but look at you in that beautiful $2,000 Loro Piana sweater and doing like posting this…
This is a regular sweater.
Um, so yeah, I… I think the relativistic point is really important, but like, you know, to us, our jobs feel incredibly important and stressful and satisfying. And if we're all just making better entertainment for each other in the future, maybe that's kind of what at least one of us is doing right now.
That's right. That's exactly. Uh, all right, let's change gears. I don't like where this is going. Let's talk about OpenAI.
A good thing about a brother is they can really call you out on your…
It's really helpful. I appreciate it. Let's go to…
Can we talk about OpenAI?
Yes. Thank you. So with OpenAI, so far we have like a consumer business. There's obviously a B2B business. There's something in hardware with Johnny IV. There's a bunch of other potential stuff that looks like it's kind of hanging around. Um, can you talk about like what's the potential complete apparatus, or what's the apparatus at least in some period of time?
Yeah, I think what… what consumers want from us eventually is an AI companion, for lack of a better word, that lives in the ether and that is helping them in all these ways through all these surfaces and all these products and that gets to know you and your goals and what you want to accomplish and your information. And sometimes you'll like type in it with inside of ChatGPT. Sometimes you'll be using a more like entertainment-focused version. Sometimes you'll be using other services that will have like integrated with our platform. Sometimes you'll be using our new device, and… but what you will have is this thing that will just be helping you get done whatever you want to get done. Like sometimes it's pushing stuff to you, sometimes you're like asking questions, sometimes it's just there like observing and getting better for the future. But that is eventually what I think it'll feel like is this is my… we don't quite have the right word for it… my AI companion is the best I can do right now.
Do you think we have like the wrong form factor with all of our stuff right now, like computing?
Yes. Uh, "wrong" is too strong of a word. I don't think we have like the optimal thing. We… we there have basically been two revolutions in computer form factors, interfaces, whatever you want to call it, that I think have really mattered. I mean, there was like stuff a long time ago, but neither… you know… we were paying attention. But in our lifetime, there's been uh… like this kind of a computer, like a keyboard, a mouse, and a monitor, which is pretty awesome and pretty general purpose. Um, and then there's been like touch devices that you carry around. And honestly, those are the big ones. Both of those had the constraints…
The lack of AI meant limitations in what could be built or drawn. With this incredible new technology, we can approach the kind of computer that exists in science fiction. It'll be the same intelligence, just in a different form factor, allowing for different usage. The form factor matters because it's with you all the time. That's crucial if it's constantly with you, full of sensors, understanding what's happening, and keeping track of many things. If you trust that a small command can achieve something complex and correctly, you can imagine very different kinds of devices.
What other components are you considering? There's how chat is used by consumers, the API used by startups, and the device aspect. What are the other key elements? The most important, yet unacknowledged, is its potential as a platform that integrates into and with everything. Seamless continuity across devices, websites, and even your car will be significant.
This opens up new possibilities for productivity and social entertainment. The ubiquity of the intelligence, with its broad implications, will be a defining factor. There are subcomponents and layers above the stack, like energy, hardware, and more. How important is this entire stack, considering its implications?
Critical. The world needs to consider this, from the electron to the ChatGPT query. There's a lot in between, and I've started calling that the AI factory. It's more like a meta factory, capable of making copies of itself. We need to address the entire supply chain. Should OpenAI handle much of it? Vertical integration has advantages, but it's not crucial for us to do everything. Partnerships can drive huge progress, minimizing risk of losing parts of the process.
On the energy side, will we consume massive amounts of energy? I hope so. Increased energy abundance has correlated with improved quality of life throughout history. I see no reason for that to stop. Are there climate concerns? I believe fusion and new fishing techniques will address this; it's the least of our worries.
How confident are you about fusion? I wouldn't say totally, but quite confident. Fusion could become a huge percentage of our energy, alongside next-gen fission. Companies like Oaklo are doing great work in solar and storage. However, at some point, even with fusion, scaling Earth's energy use too much will overheat the planet. But we have a vast solar system.
Doesn't this make space exploration more important and likely? Yes, we're going to space. It would be sad if we didn't. Should I start a rocket company? I've suggested that before. You should do rockets, social media, and many more things.
Speaking of social media, what's the situation with Meta's efforts? Meta views us as their biggest competitor, and I understand their continued attempts. Their current AI efforts haven't worked as well as hoped, but I respect their aggressiveness. This reminds me of Google's early efforts against Facebook in social media. Meta has made huge offers to our team, but our best people haven't left.
They offer massive signing bonuses and high compensation, focusing on that instead of the work and mission. I think this won't create a great culture. We aim to be the best place for this research, with a special culture. Success will result in great financial rewards for everyone. While I respect Meta, I don't think they're great at innovation. OpenAI's culture is better at repeatable innovation.
Is copying existing AI work enough, or is most innovation still ahead? It's not enough. Many at Meta are trying to copy OpenAI, even mimicking UI mistakes. This approach of following competitors rarely works. It doesn't build a culture of innovation.
How do you manage a commercial and research-oriented company simultaneously? There aren't many examples. We started as a research company and added a product company (initially badly run, now improving). We're becoming a great product company, but two and a half years ago, we were just a research lab. ChatGPT launched November 30th, 2023. It's easier to assemble a company than to do this kind of research.
Why does Meta see you as such a competitor? They likely view AI as the whole game. Someone from Meta mentioned that while others see ChatGPT as a Google replacement, Meta sees it as a Facebook replacement because people spend so much time interacting with it, and they prefer it. It's not just about time; people find doomscrolling on the internet detrimental, while ChatGPT is seen as helpful and positive.
Someone said ChatGPT is the only tech company that hasn't felt adversarial. Google shows worse search results and ads, Meta tries to manipulate users into constant scrolling, and Apple, while making great products, bombards users with notifications. ChatGPT feels helpful.
Can social media incorporate the interpersonal component and the energy of AI? A prompted feed, tailored to goals like fitness or learning current events from a neutral perspective, could be a positive alternative to outrage-baiting algorithmic feeds. This would likely result in less time spent, but it would be a more aligned experience.
I start each day with good intentions, but the day's events often lead to compromises, like drinking or scrolling on TikTok. I should work less. If I could maintain my morning self throughout the day, and if technology helped, I'd be great.
Ten years ago, I had high agency and did as I pleased. Recently, it feels like there are fewer rules. There's the Stargate project, bringing Fiji into the company, Johnny, etc. Is there a mental shift or reason for this? It's like my grandmother said; getting older means caring less about what others think. I've been through enough. It's freeing.
Are there things you hesitate about? Yes, with more resources and potential, we can do more. I'd love to build a Dyson sphere, but that's decades away. We can do more now. How do you choose among many options? I have limited bandwidth; I never wanted to run many companies. I thought I'd just be an investor.
You seem happy. I have time for hobbies. I'm glowing because my wife gave me a gua sha facial massage. It helps your lymphatic system. I like focusing on one thing well. I trust others to create great rocket companies.
Do you enjoy OpenAI, despite it being more than you bargained for? I'm grateful and lucky, and I'll miss it someday. It's satisfying and important. It's been overwhelming and stressful. I never expected this. It was supposed to be my retirement job, running a small research lab. There were many worlds where this didn't happen.
Is it heavy and important, or a playful puzzle? Both. It's potentially the most impactful work I'll ever do, but day-to-day, it's more like a puzzle. I love my team, and I find joy in the details. Some parts are stressful, but it's fun. I can talk to and think about whatever I want, except I'm a pre-trained model, reacting to a stream of things.
You can input everything, but you can't control your time. It feels more reactive. This is tough, given the pace, but it will hopefully improve as we keep hiring executives. What about your own experience?
Me? You know, obviously, I don't think any differently about you know, you on some level other than I think this is very cool. But like, you've obviously gone from like tech echo chamber famous to just like regular New York Times famous. Is there any upside, or is it all just suck?
I think there is a lot of upside to being tech famous. I think that's like the perfect level of it. Like if you're like tech famous, you can sort of get a lot of interesting stuff done. You can meet kind of anyone you want. Like it's it's just like that felt you can kind of like encourage people to do stuff. There's like a lot of good opportunities. That felt good. I mean, I think I'm not like celebrity famous. That would be really miserable, but I'm It's enough that I can't like quite have a normal life. Well, it has a different energy to it. Like you're never going to be as cool as Tom Cruise or something. But I like, you know, I can still like walk on the streets.
Um, I didn't mention this to you, but when we were in the Exploratorium last weekend in the gift shop, somebody was like, "The CEO of Chad GPT is in the building."
That's good. My son still calls Chad GPT Siri, which I think is very funny. I still We're going to talk about that.
Yeah, I talked to Siri this morning. It was very funny when you guys called this weekend and he was like, "Are you really? Are you really?"
Yeah. The inventor of Siri right here.
Speaking of kids, um, given the current trajectory you think things are on, does it make you change what you think kids should be learning or what you'll teach your own kids?
My kid did learn to roll over yesterday.
That's pretty good.
Yeah, I was very impressed. The headstrength. Incredible.
Yeah. Top top tier baby. Um, incredible. Strong.
Yeah. I don't think so. Like I I mean this stuff it's not gonna ever seem weird to him, right? Like he's just going to grow up in a world where of course computers are smarter than he is. Yeah, and he'll figure out how to use them incredibly fluently and do amazing stuff, and that will just seem like Yeah.
Yeah. That's good. Hopefully that's how it plays out.
When you go back to like YC after this experience, are you back and are you like, oh, this is like, you know, cuz obviously you ran YC, and so you had that whole chapter of your life. When you go back now, are you like how do you experience it? Is it like, oh, this is like a quaint thing, you know, all these?
No, I love YC. I I I think like I don't mean that in a negative way. I mean, in just sort of like a simpler, simpler times kind of way. There's some nostalgia for that.
Yeah. Yeah. It's really like fun to always like go back and talk to the batch. Uh, but yeah, it hits the nostalgia hard. I just feel like it's like the most pure component of Silicon Valley is what I've seen. I totally strong. You know what I mean?
100%. It's just so happy. It is earnest and positive and energetic and happy and like clearly it works super well.
Yeah. All right. Well, this was a blast. Mom, hope you were watching this. This was very fun.
Thanks, Tim. Thank you, Jack.