Sam Altman CEO of OpenAI: ChatGPT, Future of AI, AGI, and China

Hi everyone and welcome to our podcast In Good company. I'm Nicolai Tangen, the CEO of the Norwegian Wealth Fund. Today, we are releasing a very special bonus episode with the one and only Sam Altman, CEO of OpenAI. Now. This conversation just blew my mind away. You have to listen in.

Thanks a lot for having me, it's great to be here.

How does it feel to spearhead this revolution?

It's definitely a little surreal. It's like a very exciting moment in the history of technology. And to get to work with the people who are creating. This is a great honor. I can't imagine anything more exciting to be doing.

No, I can't imagine.

It's definitely a lot.

I can see that now. Big picture What's the vision of the world where humans and AI co-exist?

Well, you know, one thing that we believe is that you have to answer that question empirically. There's been a lot of philosophizing about it for a long time. Very smart people have had very strong opinions. I think they've all been wrong, and it's just a question of how wrong. The course that a technology takes is difficult to predict in advance.

How far into the future can you see now?

The next few years seem pretty clear to us. We kind of know where these models are going to go. We have a roadmap we're very excited about. We can imagine both the science and technology, but also the product a few years out.

Lot smarter in two years than we are today. What kind of holy shit moments have you had lately?

Well, remember that we've been thinking about this and playing around with this technology for a long time, so the world has had to catch up very quickly. But we have less holy shit moments because we've been expecting this and we've been building it for a while. It doesn't feel as.

Discontinuous to us.

The biggest ones have not been about new technology or new models, but about the breadth of use cases the world is finding to do this. So the holy shit moments have not been like, Oh, now the model can do this now we figured out that because, again, somewhat expected that. But seeing how much people are coming to rely on these models to do their work. In their current form, which is very imperfect and broken, we're the first to say these models are still not very good.

Has been big and some examples of new things you've seen, new user cases.

Applications.

And what about things like brain implants and getting it to help with speech and so on?

Which we just saw recently.

Now, a few years ago, nobody had heard about OpenAi, now everybody's heard about it. You are one of the most famous people on earth, so how many people are you at OpeNAI now?

500 or so.

And what do these 500 people actually do?

It's a mix, so there's a large crew that's just doing the research, trying to figure out how we get from the model we have today, which is very far from an AGI to an AGI and all of the pieces that have to come together there. So scaling the models up, coming up with new methods.

And so your priorities as a C.E.O now? how do you spend your time?

I kind of think about the buckets of what we have to do in research, product and compute on the technical side. And that's sort of the work that I think I enjoy the most and where I can contribute the most. And then I spend some of my time on policy and sort of social impact issues for lack of a better word, and then the other things I spend less time on. But we have great people that run the other functions.

Now, your mission has been to ensure that artificial, or the general intelligence, benefits all of humanity. What's the biggest challenge to this, you think?

A couple of thoughts there, one I'm reasonably optimistic about solving the technical alignment problem. We still have a lot of work to do, but I feel better and better over time, not worse and worse. The social part of that problem. How do we decide whose values we align to?

How far out is true AGI?

I don't know how to put a number on it. I also think we're getting close enough that the definition really matters, and people mean very different things when they say it. But I would say that I expect by the end of this decade for us to have extremely powerful systems that change the way we currently think about the world.

And you say we've got different definitions. What is your definition of general intelligence?

So one that matters, there's the open AI official definitions, and then there's one that's very important to me personally. When we have a system that can figure out new scientific knowledge that humans on their own could not, I would call that an AGI.

And that you think we may have by the end of this decade?

Well, I kind of tried to soften that a little bit, just by saying, we'll have systems that really change the way the world works. The new science may take a little bit longer, or maybe not.

What's the endgame here? Are we just all of us going to work a lot less?

I think we'll all work differently, I think many of us will still work very hard, but differently every technological revolution. People say that we're just going to do less work in the future. And we just find that we want a higher standard of living and new and different things, and also that we find new kinds of work we really enjoy.

I believe you just traveled the world and met with a lot of people and users. What was your main takeaway?

The level of excitement about the future and what this technology is going to do for people around the world in super different cultures and super different contexts.

Any difference between geographies or industries?

Yeah, in the developing world, people are just focused on what this can do economically. Right now, in the more developed world, there's much more of a conversation about what the downsides are going to be and how this is going to disrupt things. There's still excitement, but it's tempered more by fear. That was a striking difference.

Do you think it will lift up the poor part of the world?

Yeah, I really do. I think it's going to make everybody richer, but I think it positively impacts poor people the most. I think this is true for most kinds of technology.

What would potentially prevent this from happening?

Well, we could be wrong about the trajectory that technology is on. I think we are on a very smooth, exponential curve that has much, much further to go. But we could be missing something, we could be drinking our own kool-aid, we could hit a brick wall soon. I don't think we're going to.

What is it doing to the global balance of power?

I don't know how that's going to shift. I'm not sure anyone does, but I certainly don't think that's something that I'm particularly well qualified to weigh in on.

It just seems like it's so keen out to the weapon race, the medical race, the self-driving vehicle race, just all these races.

It's also available pretty broadly. One of the things that we think is important is that we make GPT-4 extremely widely available. Even if that means people are going to use it for things that we might not always feel are the best things to do with it. We have a goal of globally democratizing this technology. As far as we know, GPT-4 is the most capable model in the world right now.

The right chips, right, they are not available.

People, I would argue that's even more of a constraint. There are people who can go copy GPT-4, right? We did it fine. Once people know how to do it and know that it's possible, a lot of people will replicate it.

Here we don't, maybe someone does here. Do you think there's a chance that? Well, like they did with weapons that just suddenly bang, they had the supersonic rockets, we didn't even know they.

Existed, right? Could that happen? Yeah, totally, it could.

So Marc Andreessen, for instance, he thinks we should stuff it into everything.

Stuff it into everything means just put it everywhere. Yeah, that's happening, and I think that's great.

Some examples if you had to reveal something.

I mean, like, you know, car makers are putting it into cars. And I was like, all right, that sounds like a gimmick. And then I got to try a demo of it and I was like, Wow. Being able to just talk to my car and control it in a sophisticated way, entirely by voice actually totally changes my experience of how I like, use a car in a way that I would not have believed was so powerful.

So, for instance, use it in a car. What do you say?

This is probably where I don't want to like reveal a partner's plans. But you can imagine a lot of things that you might say, like, the basic stuff is easy. Like, you know, I need to go here and I'd like to listen to this music. And also, can you make it colder?

Sounds good. Do you depend on newer and even more powerful chips than what we have now? I mean, how much quicker do you? How much more complex do chips need to be than H100? Or the latest things from Nvidia?

Yeah, of course. Like, there's the ways that we can keep making these models better are we can come up with better algorithms, or just more efficient implementations, or both. We can have better chips and we can have more of them, and we plan to do all three things and they multiply.

Together.

They will end up with profits, I wouldn't say the profits. I think there's many people who are going to, like, share this massive economic boon.

How much does it cost to train these models? I mean, how much have you spent on pre-training models?

We don't really talk about exact numbers, but, like, quite a lot.

And what's the challenge of spending so much money pre-training? And then it lasts for a relatively short period of time? In a way, you have to depreciate the whole investment in order, because you need to invest more in the next generation.

I, That's true. I don't think there are going to be as many massive pre-trained models in the world as people think. I think there will be a handful and then a lot of people are going to fine-tune.

On top of that, or whatever.

Part of it that I think is important is like, you know, when we did GPT-4, we did. We produced this artifact and people use it, and it generates all this economic value. And you're right, that does depreciate fast. But in the process of that, we learn so much about how to go. We push the frontier of research so far forward, and we learn so much. That'll be critical to us being able to go do GPT-5 someday or whatever. That it's like, you're not just depreciating the Capex one time for the model, you have generated a huge amount of new IP to help you keep making better models.

So the way you read the competitive landscape now, how? what does it look like?

I mean, there are going to be many people making great models. We'll be one of them. We'll like, contribute our AGI to the world, to society, among others. And I think that's fine, and you know, we'll all.

So normally in the software business, you have something which is very cheap, where you ship a lot of it, or something which is very expensive and you don't ship so much. Here, you could potentially ship something, and I can see you smiling here. here you can potentially....

We ship a lot of various types of things.

Exactly so, tell us how is this going to work?

You know, I'll tell you, one of the most fun things about this job is. We are past the point as a company. I am past the point as like a CEO running this company. Where there's like a roadmap to follow. We're just doing a bunch of things that are like, outside of the standard. Silicon Valley received wisdom. And so we get to just say, well, we're going to figure it out and we're going to try things. And if we got it wrong, like, who cares? There was no like, it's not like we like, screwed up something that was already figured out.

I'm sorry, product company, you mean, Microsoft.

No, no, I mean, like, having to figure out how to ship the API in Chatgpt, yeah.

Is the usage of ChatgPt decelerating?

Nope.

Doink up.

I mean, at a high level, they build us computers, we train models, and then we both use them.

Are you have you? are your goals aligned?

Yeah, they really are one of. I mean, there's like, there's, of course, areas where we are not perfectly aligned. And like, I don't like any partnership in life or business or whatever. I won't pretend it's perfect, but it is very good. And we are aligned at the highest levels, which is really important, and the misalignments that come up at the sort of lower levels. Once in a while, you know, like,....

And I mean, together with you, in terms of self-regulating this space, what can this type of thing be self-regulated?

Not entirely. I think it needs to start that way, and I think that's also kind of like how you figure out a better answer. But, like, governments are going to have to do their own thing here, and you know, we can provide input to that, but we don't. We're not like the elected decision-makers of society, and we're very aware of that.

And what can governments do?

Anything they want, and I think people forget this. Like, governments have quite a lot of power, they just have to decide to use it.

Yeah, but I mean, so let's say, now Europe decides that they're going to regulate you really harshly. I mean, are you just going to say goodbye Europe? No.

Possibly I don't think that's what's going to happen like. I think we have a very productive conversation. I think Europe will regulate AI, but reasonably not very harshly.

And what is a reasonable regulation? What is that level?

I think there's many ways it could....

I agree, I agree. Do you think we'll get global regulation? Is there any shape or form that can happen?

I think we're going to get it for only the most powerful systems. So, you know, I think like individual countries or blocks of countries are not going to give up their right to self-determine. For like, you know, what can a model say and not say? And how do we think about the free speech rules and whatever?

But given how embedded it now is in everything, as we spoke about, you know, weapons. Your car, you're sitting in your car and it's like, super cool, and it's cold and hot, and music and this and that. And you know, and you're a Chinese car company and you want to outcompete the Americans. Why do you want to have a regulation on this?

Well, GPT-4, I don't think needs global regulation, nor should it have it. I'm talking about, like, what happens when you get to GPT-10? And it is, you know, smarter than all of humans put together.

And that's where you think we'll get it.

That's when I think we'll get it.

When you have the cost of intelligence coming down so dramatically like it is now, what is it going to do to productivity in the world?

I mean, it's supposed to go up a lot, right?

That's what theory tells us.

That's what I think.

So I've told everybody in our company that, hey, we should improve our productivity by 10 over the next 12 months. All of us, and that's, and you know how I got the number?

Did you ask to GPT?

No, I just took it, I just took it straight out of the air.

Underambitious.

Underambitious? What, what should, what should productivity increase by?

How do you? how do you measure?

The stuff we do.

How much of your company writes code?

If 15, well, people in in technology, probably 15 20 of us.

I think an overall goal of, like, you know, 20 productivity increase in a 12 month period is appropriately ambitious. Given the tool and given the tools that we will launch over the next 12 months.

Okay, sounds like I should up the game here a bit.

I think so.

Yeah, I'll just tell everybody you told me to, so that's fine.

It's better to set a goal that is like, slightly too ambitious than significantly underambitious, in my opinion.

Yeah.

I couldn't come up with any reasonable explanation of why that should be the case.

You say that most people overestimate risk and underestimate reward. what do you mean by that?

Um, you know, there's a lot of people that don't go start the company or take the job they want to take, or try a product idea because they think it's too risky. And then if you really ask them like, all right, can we unpack that? Can you explain what, what the risk is and what's going to go wrong? It's like, Well, the company might fail, okay?

Talking about staying in your job, what? So the leaders and the CEOs, so you know how? How is AI going to change the way leaders need to act and behave?

Well, hopefully, it's going to, like, do my job, you know? Hopefully the first thing we do with AGI is let it run open. AI, I can, you know, go sit on the beach, that'd be great.

How do you develop the people in your company? How do you develop your leaders?

Um, I think developing leaders tend to fail at the same set of things most of the time, you know, they don't. They don't spend enough of their time hiring talent and developing their own teams. They don't spend enough of their time articulating and communicating the vision of their team. They don't spend enough of their time thinking strategically because they get bogged down in the details. And so when I like, put a new person in a very senior role, which I always try to do. With promotions, I mean, I'm willing to hire externally, but I'd always, always rather promote internally.

What's the most common way leaders grew up?

Uh, failing to.

Uh, what, what is your delegation process?

Two things number one, high quality people, number two, setting the training wheels at the right height and increasing them over time. As people learn more and I build up more trust.

Is that the way to manage geniuses?

Um, they get, uh, uh, researchers, that's a different thing. I was like, talking about how to like executives that run the thing.

Okay, what about researchers? what about the geniuses?

Um explain.

Um, do you think companies generally are too detailed in the remit they give the teams?

Yes, I mean, at least for our kind of thing, I think, uh, managing.

I'm slightly surprised you say that there was no innovation culture in Silicon Valley, because that's, uh, a bit contrary to, uh, to what I thought.

So, yeah, there's a product innovation culture, for sure, a good one, but like, I mean, again, I hate to say this because it sounds so arrogant. But like before Open AI, what was the last really great scientific breakthrough that came out of a Silicon Valley company?

And and why did, and why did that happen, why, what happened there?

Well, we got a little lucky.

No, I don't mean why, I'm saying, why did? Why did this culture disappear in Silicon Valley?

I have spent so much time reflecting on that question.

Now you, you had some, um, what should we say? Your your co-founders are pretty pretty into big, uh big hairy goals, right?

Yeah, I mean, we're trying to make AGI, I think that's the biggest area is going in the world.

So not so many companies have those kinds of co-founders and people who with that kind of track record, and you know that that type of talent magnet, uh, funding capabilities and so on. Do you? How important was that?

You mean Elon by this, right?

Yeah, yeah, and uh, you know, and some of the other people, you, you, you, you, you, you, you, you you.

You, well, there were. So there's six co-founders, Uh, you want to meet Greg and Ilya and John and Vujicic? And, you know, Elon was definitely a talent magnet and attention magnet, for sure. And also, just like, has some real superpowers that were super helpful to us in those early days. Aside from all of those things, you know, contributed in ways that we're very grateful for. But the rest of us were like, pretty, pretty unknown. And I mean, maybe I was like, somewhat known in technology circles, cause I was running Y combinator, but not not like a not in a major way.

What is your superpower?

I think I'm good at thinking very long-term and not being sort of constrained in, like, common, common wisdom.

Yeah.

You said in 2000, I think I'm good at evaluating talent. That was like a really helpful thing to learn from Y Combinator.

You said in 2016 that's long-term thinking is a competitive advantage, because almost no one does it.

Yeah, I mean, when we started opening AI and said we're going to build AGI, everybody was like, That's insane. A, it's 50 years away and b. It's like, you know, the wrong thing to even be thinking about. You should be thinking about this, how to improve this one thing this year. And you know, also, this is like, unethical to even say you're working on it because it's like such a science fiction and you're going to lead to another AI winter.

You said you also good at assessing talent. how do you do it?

I don't know, I don't, I can't like, I have a lot of practice, so I've got like a, but I don't have like words for it.

Why is Europe so behind generally when it comes to innovation and innovative culture?

I'd ask you that I don't know why, is it?

Well, is it first of all? Like, Well, I guess, I guess it is.

It's certainly behind, it's certainly very behind in, like hyperscale software companies, there's no question there.

It's big, big fear of failure, it's a cultural thing. Um, it's, there's a lot of, there's a there are a lot of things going into that cocktail.

I think the the the fear of failure thing, um, and and the kind of the like the cultural environment or backdrop there is is huge, no doubt. Uh, the, you know, we funded a lot of European people at YC. And a thing they would always say is like, they cannot get used to the fact that in Silicon Valley, failure is tolerated. Yeah.

What's the biggest failure so far?

Uh, well, I mean, monetarily wise, I've made a lot of big investments that have gone to total like, just, you know, zero. Like crater in the ground. But in terms of, like, time and psychological impact on me. I did a startup from when I was 19 to 26, worked unbelievably hard to consume my life and failed at that. And that was like, quite painful and quite demoralizing. And it's like, you know, you learn to get back up after stuff like that, but it's hard.

How'd you get back up?

Um, I mean. One of the key insights for me was realizing that, although I thought this was like terribly embarrassing and shameful, uh, no one but me spent much time thinking about it.

Who do you ask for advice like personally?

My strategy is not to just have like one person that I go to with everything, and a lot of people do that, you know, they have like one mentor that they go to. For every big decision, but my strategy is to talk to a ton of different people when I'm facing a big decision.

So now outside, AI, Um, what are you the most excited about?

Fusion I think we're going to get fusion to work very soon and I think my model, if you boil everything down to get to abundance in the world. The two biggest, most important things are bringing the cost of intelligence way down and bringing the cost and amount of energy way down. And I think AI is best way to do the former, and fusion is the best way to do the latter. And, you know, in a world where we look at energy, that's like less than a penny per kilowatt hour.

Do you think it's going to solve the climate problem?

Yes.

What about crypto?

Uh, I am excited for the vision of crypto, and it has so far failed to deliver on that promise.

But you have plans.

It's it's not something I'm spending that much time like. OpenAI has taken over my whole life, so I can have a lot of plans about OpenAI. And there's other projects that I've invested in or helped start that. I feel bad because I don't have much time to offer them anymore, but they're all run by super capable people and I assume they'll figure it out.

What do you read?

Um, the thing that has unfortunately gone.

What's the one book young people should read?

That's a great question. Picking one is really hard.

Good, now we are, we are, uh, fast forwarding.

Oh, you know, what can I? Can I? Actually? I do have. This is not the one for every young person. But I wish a lot more people would read the beginning of Infinity early on in their.

The beginning of Infinity?

Beginning of Infinity, Why, uh?

Oh, it doesn't matter, we'll find it.

I think it's the most inspiring. You can do anything, you can solve any problem, and it's important to go off and do that. It's a very like. I felt. It was like a very expansive book of the way I thought about the world.

Well, Sam, I think that's a very suitable place to, uh, to go in for landing now. Last one. So, um, fast forward a couple of decades, um. People sit down and reflect on Sam Altman's impact on the tech world and society. What, what do you hope? What do you hope? They'll say? What do you? What do you hope your legacy will be?

You know, I'll think about that when I'm like, at the end of my career, like right now, I. My days are spent like, trying to figure out why this executive is mad at this one and why this product is delayed. And like, why our network on our, you know, big new training computer is not working? And who screwed that up and how to fix it? And it's like, very caught up in the like, annoying tactical problems. Uh, there is no room to think about legacy, we're just trying to go off and like, build this thing.

Fantastic. Well, um, good luck with that. It's been an absolutely fantastic conversation and, uh, all the best of luck, and, uh, go get them.

Yeah, great talking to you, thank you for, thank you for having me.

Very cool.

Wow, that was great.