Eric Schmidt: Singularity's Arrival, the 92-Gigawatt Problem, and Recursive Self-Improvement Timelines | 241

0:00We're living through a historic moment right now. The next thing that's really interesting and terrifying also is recursive self-improvement, but we don't have it yet. What we do have, I keep asking my friends, when does the asymptote arrive and when does the curve slow down? It is actually true that there is a limit to our craziness. We have not found it yet. And that's the great thing, frankly, about America. The American competitor, not enemy, but competitor is China. They have lots of money. They're very, very smart. Their work ethic is equal or stronger than ours, and they dominate key industries.

0:38But at the moment, it sure looks to me like the robotic hardware of China is the winner. I don't want to lose the robotic revolution, in my view, the way we lost the electric vehicle revolution, at least on the low end. It's possible, but it requires... Now that's the moonshot, ladies and gentlemen.

1:02Eric, do you remember the first time we met? Yes, Larry Page introduced me to you because he was on your board. Yes, yeah. So I got a call from Eric out of the blue, which was a great honor, and said, Larry says I should meet you. When are you going to be down here or up here in San Francisco? I was in L.A., and I said, how about tomorrow? Typical Peter. And I remember something which was so, I love the story. We sat down to lunch at Charlie's Cafe. And of course I'm running a non-profit and my mission is always raising capital for a non-profit. And so we sit down to lunch and before we get started you said, Peter, so what is your highest level of giving of membership?

1:48And I said, well Eric, it's our vision circle for$2.5 million. and you go, okay, I'm in. Now let's have a conversation. I'll buy them all. It was crazy. Oh, there you go. But, you know, somehow you have, the reason when we spoke, the reason I wanted to come here is this has become the epicenter of the abundance movement. And the abundance movement is correct. That's the important thing. Thank you. Yeah.

2:21I just want to thank you for all the support you've given myself, XPRIZE, all these years. So grateful for that. So let's start with a question that I'd love to hear you expand on, which is we're living through a historic moment right now, right? Could you define the moment we're in and give us sort of a state of the union of what's going on in AI? we're 10 or 15 % into the impacts of this. And you can see it. You can feel it. And some of it will happen. Some of it will take longer. So, for example, hardware takes longer than software. So do robots take longer than digital systems on traditional hardware?

3:09Things like that. We've not... The next thing that's really interesting and terrifying also is recursive self-improvement. It's not happening yet. And so it's easy to convince yourself that you're going to have human agents, sorry, computer agents that are human-like completely within a year or two. We don't have the science for that yet. People are working on it. I can describe how I think it'll play out. But we don't have it yet. What we do have is reasoning systems that are perfect partners for human beings, for good and bad, right? And that has a lot of implications. So if we stop today, which we're not, and it's not stoppable or controllable by any government or any single individual or corporation, we would still have advanced humanity because of these reasoning agents.

3:59How fast do you imagine this is going to accelerate?

4:05There's a thing which I call the San Francisco consensus. And the reason I call it is because everyone in San Francisco believes this. Everyone I know, anyway. which is that it's easy to understand. This is the year of agents, which we can discuss, why agents will take over everything this year. During this year, the scaling of the use of agents and reasoning will sort of grow at this enormous rate. Everybody's out of hardware, everyone's out of electricity. It's a real boom, right? It's like the biggest boom I've seen, and I've been through three or four of these in my career.

4:41In this thinking, once you have recursive self-improvement, where the system can begin to improve itself, you have intelligence learning on its own. And it will, in this argument, it will learn faster than we can, because we're biologically limited. And the way this is expressed in San Francisco, and I'll give a simple example, you have a tech company with a thousand fantastic AI researchers. so one day they turn on AI research that is an AI research agent well how many AI research agents do you have well as many as you're limited by electricity right you don't have to feed them they don't need housing there's no more housing in San Francisco you know all that kind of stuff but you don't have those problems you don't have an HR department for them if you will and you don't have to pay them you just have to feed them electricity so how many could you have well, maybe a million of these agents.

5:35Now in AI, the way you determine that you've made progress is you have clear metrics that the reasoning or testing or whatever, the evaluation framework is better, right? So that's what happens. So in that scenario, the slope goes like this because you're already at this slope, then you add more people, then you get the agents and you go like this, and this is essentially a super intelligence moment. The belief in San Francisco is this occurs within two to three years. The evidence in favor goes something like this. Claude Code came out a couple months ago, the latest one, Opus, whatever it is.

6:194.6. Yes, thank you. And everyone I know in the Bay Area that's doing software says it was 80-20, now it's 20-80. the best analysis I can come up with is it's not the Claude code part it's that the underlying LLM can produce more reasoning over time better quality tokens over time it's a deeper thinker and all the labs now are competing for that this is not just the size of the context we're into it's actually the reasoning skill and the length of which it can think it can just think longer and produce more stuff I watched this stuff when I was, I moved to the Bay Area when I was 21 and I was a programmer in high school way back when and I was a pretty good programmer and I watched what it does and I go, my God, I'm over there's not a thing that I could do that it cannot do so when they wrote a C compiler in Rust I thought, it's over like declare success so I think part of this part of this is because the people who are building it are also seeing the diminution of their own skill.

7:35They're being forced to go from programmers, which is what I'm very proud to have been, to being the director of a programming system. And the most likely scenario, by the way, there's a lot of implications for this. One is that it's always been true, speaking as your local arrogant programmer, that the very top programmers were worth 10 times more than the ones right below. There's something special about the mathematical reasoning skills of programmers. Those people will become more valuable, not less valuable, because these systems need to be controlled by humans at the moment. Those people will be capable of grasping the parallelization and the activities of this.

8:15It also means that you're going to have, in my view, a relatively small number of very large companies. Yeah, yeah. And this is a big deal. Yeah. And a very large number of very small companies because you don't need as many people. And you're watching that play out this month. I mean, this all happened in the last three. I was in one startup I'm involved with. I was talking to the programmer who's a perfectly brilliant young man. And I said, well, what's the truth? He said, well, here's what I do. He's working on UIs of various kinds. And he said, I write the spec of what I want, and then I write a test function, an evaluation function, and then I turn it on.

8:57I said, what time? And he goes, seven o 'clock in the evening. And I go, okay, what do you then do? Well, he has dinner with his wife, and he goes to sleep. And I said, do you wake up? He said, no, I sleep very well. When does it finish? Oh, four in the morning. And then he gets up, has breakfast, does whatever he does, and then he sees what's been invented. I mean, it's mind-boggling. And the stupid example I used with this young man, this is what the power of these systems are. If you can define the evaluation function and you can let it run, and if you have enough hardware, you're inventing worlds.

9:36I mean, this stuff would have taken me six months and 10 programmers at Google to do the same thing. This poor guy's sleeping. It's so funny you say that because I was literally backstage. They said, Eric Schmidt's coming now and I have my lid open on my Mac and I'm trying to get the jobs onto the cloud so I can close the lid because if you close the lid, it'll break the jobs. So I've got these six concurrent quad 4.6. You can leave it open on the, you know, it's like important what you're doing. Don't interrupt for me. You're important too, you know. No, it's crazy because when we got together in Davos just, what, two months ago, it was in this kind of autocomplete mode.

10:12You'd write the code and then it would help you get it done. You're about 10 times more efficient, but you're still babysitting it. Now, literally, it's working right now. When I get off stage, it will have solved six problems that I launched. And I appreciate the excitement in the industry, but I can tell you when I was, I used to work on BSD, I basically worked at Berkeley on Unix at Bell Labs and at BSD Unix. And programmers invent what they need. And so we invented the first email system and the first messaging system. And nobody thought about it. It was like, well, we just need this thing.

10:45Yeah. So one key thing to understand about digital intelligence is the first inventors are the people who are solving problems of themselves, who are programmers, right? You shouldn't be surprised by this. You should have expected this. The other thing that's interesting about programming is it's both scale-free, which means there's no particular limitations except electricity. You don't need a lot of data, and you already have GitHub and the equivalents. And it's also a fairly limited language set. So the number of language components, if you will, compared to human language is less. Smaller language, clear objective function, all you need is electricity.

11:21Now, how far can this go? It'll get to the point where you don't have the ability to do completely new things. Isn't it really quaint and crazy to think that we can sit here and say, yeah, I wrote a ton of code when I was younger. No one will ever do that again after the end of this year. It'll be like riding a horse. It'll be like those quaint skills that we all used to have. I do have a proposal for universities. Those of you who are associated with universities, you should stop everything else you're doing in the university right now and design a course for freshmen, men and women, starting in September, which is a prompt engineering class.

11:58Why university? Why not high school? God, you're so aggressive, Peter. Let's start with universities. You can improve my idea. I thought 18-year-olds would be young enough. Maybe you think it's younger. Here's the most important thing. Spend a quarter or a semester. The first thing they learn in university is how to use these tools. Universities are completely opposed to my idea, as usual. Because it violates every one of their tenets. But if you think about the student, and I mean every student, liberal arts, math, whatever, this platform will be the expression platform for their art, their music, their writing, and so forth.

12:40Why wouldn't you teach them immediately, Peter, improve my proposal? No, I just feel like we're, that AI is going to impact every student in high school today. And that they're living an unnatural life by not engaging with it. And when they hit universities for those that still exist... Well, plus your kids are that age, and so they're literally right now doing exactly what you're describing. People here who have teenagers, you know what I'm talking about, because they're all in it already. So I think that's an improvement to my argument. There's a problem of age restriction. You really have to think about vulnerable teenagers with this technology.

13:23I did some analysis of where the real problems are with this stuff. A simple summary is that at some point there will be jobs impact from this stuff. We're seeing it in software and we're seeing it in certain customer service industries, not across the board. At some point that will happen, that's an issue. Another one is how do we as a country maintain our moral values while we're also racing against China. Another one is the impact on young people. It is not okay for 13-year-olds to be committing suicide because of an LLM. It's just not okay. It needs to be addressed. It needs to be addressed right now.

14:06For sure. And there's all sorts of other issues. The other one I came up with was in agent orchestration, agents can be combined. I've always been worried that when you put the agents together, especially if they're from non-compatible vendors, you get unpredictable effects. So these are problems to be solved. So we herald the future and we solve the problems that it brought. So we're going to talk about China and government and jobs. But before we do that, I want to say, I'm in the savor the moments kind of mode right now, because I feel like the world a year from today will be nothing like the world today.

14:41And everything we're doing right now, I've enjoyed so much for so long. And I just want to savor the moment. But reminisce for one minute about the fact that while you were running Google, the transformer got invented there, the TPU got invented there, Dennis Hassabis solved protein folding, which is now universally used. Does the work in, I think it does the work in an hour that used to take a PhD student four years. Their thesis. Four years, yeah. It's like 300 million times more efficient. All of that, and all the diaspora from that, all the people working in the field in San Francisco, as you mentioned, they all were your people.

15:19And so you were there at the creation of everything we're experiencing right now. Is there anything that profoundly strikes you about that moment? Did you even realize at the time? No, I think when you're making history, you typically don't know it. I give a lot of credit to Larry and Sergey. because they were ahead of me. I'm an operating CEO, and they pushed, pushed, and pushed for excellence. And I'll give you an example. In the early years of Google, Larry would, my favorite interaction was one day I said, we need to hire some people doing Java. And Larry and Sergey said, this is the stupidest idea we have ever heard.

15:58And I could never tell with them whether they were being serious or not, or they were just joking with me. But their argument was that real programmers were programming one level lower. Today, Google has many thousands of things. But they were so precise and so driven to excellence in technology that I could not fool them. I couldn't market around them. I needed to have the technical expert. And they say, oh, that's boring. Don't do that. That's another one of your ideas. We want a new idea. And I give them a lot of credit for it. But what about the TPU in particular? Because I didn't even hear about it until much later.

16:35And it takes years to design and build your own internal chips, and now it's about to explode. I don't know how much is public or not, but it's just a monster. TPU version 1 was essentially a matrix multiplier of a particular kind. When they went to version 2, they changed the algorithm in a complicated way. And it's particularly good for inference. Whether it's brilliance or just luck, those decisions made 10 years ago set up the TPU as the perfect inference engine. And for everybody's benefit, inference is what the reasoning steps that I'm describing are. So Google is particularly well positioned.

17:12As you know, NVIDIA purchased Grok for the reason of getting that integrated. Yeah, trying to catch up to what you thought of 10 years ago. And what's interesting about NVIDIA, if you look at them, I was looking at the Rubin architecture. They managed to do, NVIDIA managed to do what Intel could never do. Intel could never get control of the complete server architecture. And they tried. NVIDIA has managed to build real supercomputers that you can really buy with enough time and money and so forth, and it will really be delivered to your...and it just do the whole thing. These are major industrial achievements, and that's why both companies will do incredibly well for that reason.

17:49Eric, in the AI exponential growth right now, talk to me about where the constraints are. You were in Congress talking about energy, chips, people, capital. Where are the constraints right now? It's interesting. I started a data center company with my friends.

18:11In my testimony, I said there was an estimated 92 gigawatts shortage of power in America between now and 2030. And by reference, a nuclear power plant is about 1.5 gigawatts. So that's about 60 nuclear plants, and we're doing essentially zero or one, right, depending on how you count. So I got interested in the question of what's the real resource constraint in America, and it's electricity. We have the universities, we have the smart people, we have the economics, and we have these amazing finance people who will give all of us billions and billions of dollars on a wing and a prayer. I mean, there's no country where the finance people are sufficiently crazy to do that.

18:55It's not true in China. It's certainly not true in Europe. And these guys are incredibly jealous of the American financial system. So I always start, thank you to the finance people for funding our dreams, whether they work or not. Thank you.

19:14So there's usually a retort at this point where people say, well, the algorithms will ultimately require less energy. I am sure that that is true. There is this property that as the power of the hardware goes up, as the algorithms become more efficient, you don't need less power. You need even more power and even more computers because we discover new uses. Jevons Paradox. It's called Jevons Paradox. And so I think that because humans have trouble with exponentials, everyone says, oh, well, in six to nine months, it'll be the bubble and so forth and so on. There's no sign of this. We've been, I and a team have been working on this for years, that the ultimate essentially scaling laws are not done yet.

20:04I keep asking my friends, when does the asymptote arrive and when does the curve slow down? We've not seen it yet. There will be one, right? It is actually true that there is a limit to our craziness. We have not found it yet. And we're running to the wall. And that's the great thing, frankly, about America. You think it's a limit to the capital or a limit to, if you just add more and more and more scale and parameters, something just doesn't work? Well, the first question is, is there a limit to the capital available? A gigawatt of power corresponds to about$50 billion of hardware, software, data centers.

20:47It's on the order of, depending on what numbers you use. So 100 gigawatts is, do the math. Can we raise$5 trillion over five years? Yeah. That's the strength of America. Could we double that? You know, we're getting... The data center build-out is 1 % of the GDP of America growth right now. Yeah, we're back to Apollo program. Thank you. And the current estimate of electricity use in America is that 10 % of the electricity in the United States will be used in the data centers. And these data centers, these are not the data centers I used to build at Google, which seemed tiny by comparison. They were immense at the time.

21:30The standard data center that's being built is on the order of 400 megawatts. These things are plus or minus, and they're about half a mile long and about 500 feet wide. And they're essentially air flow machines. Take the air, send the air out, cool it in the middle using typically air cooling, and then they have a water system inside to keep the chips. Using NVIDIA as an example, the chips are water-cooled, and also the HBM3E and now the four memories put out so much heat that they have to be water-cooled. The chips are two kilowatts. I mean, this is insane. These things will kill you. You want to hear something truly astounding and funny in hindsight?

22:08When you bought DeepMind, everybody thought it was like 800 million bucks or something. 600 million. 600 million bargain. Everyone thought, why on earth would you waste$600 million on this zero revenue AI? All it does is play Go. And then years later, it came out that the acquisition paid for itself just by controlling the air conditioning more efficiently in the data centers. The entire acquisition price was paid off. And that became the AI that's changing the world. And that one, the credit of that one actually goes to Larry Page. Larry had studied AI when he was a Stanford graduate student.

22:41and we always deferred to him in this thing. And he said, this is the best team. And I think Elon and Larry competed over it. There was some complicated kerfuffle there. And so Jeff Dean, who's the chief scientist, went over and then he and I basically finished the deal. And I still remember, you know, they're in one floor, you know, these sort of British people and led by a sort of Greek, British person, Demis. and they were smart, but Google's full of other smart people. To give you an example, in 2016, so Demis announces that we're going to win the game of Go, I figured. And by the way, at this point, they're a separate group.

23:25We've let them alone because they have to grow and figure out what they're doing and all that. And this is the patience of capital. We could let them do that. We didn't require that they do anything. And so he goes, I'm going to go. I I said, well, I'm going to come too. And I said, okay. So I fly to Korea, and it's all one floor, and I go and I meet the team that's winning the game. And, of course, there's all of these Koreans, and they're all very excited about this because they know they're going to beat the computer. And so the Koreans are in one room, I'm in another. I go to the Korean room, and they're all saying, we're going to beat the crap out of this Google room.

24:01And I go into the Google room, and it's very quiet, and there's a monitor. And there's basically what I now understand is an RL prediction mechanism of whether we're winning or not. Yeah, cool. Okay, and it starts at 50-50. So I watch, I go listen to the Koreans talk for a while, and then I go watch the screen, and it goes to 51%. Okay, and then it goes to 52%. And then David, who is the guy who is the architect, says, well, we just plan for it to get to infinity. I said, okay. So basically, it's the abundance theory. So it's just, boom. And everyone is, all the humans are crushed. They're all crying.

24:45And deep-minded people say, yeah, yeah. They were supposed to. Yeah, welcome. And then I understood the genius of the deep-minded people. And you can see this today with Gemini. Gemini 3 is probably the broadest of the non-Chinese systems in terms of its depth. because it's multilingual, multimodal, and so forth. So many moments in your life are just turning points in history. And I don't know if you realize them in the moment, but that was one of the last moments where humans used to look for challenges where the computer could try to catch up, like chess. And then I think Go was the end point.

25:21And we knew that, by the way. We understood that Go, the game of Go, was sort of incomputable by normal algorithms. There was lots of math that said you couldn't solve it. And so it was the combination, they came up with a two-tree model with two different RL trees. So one of the other things I learned about these things is it's not just write me a Go program that will win the game. You actually have to understand the game and so forth. So they took the same team. They got bored with Go having won. So then they took the same team and had them work on protein folding. And in protein folding, they took a whole bunch of protein scientists, which I know.