Eric Schmidt 谈机器人竞赛、奇点时间线与能源短缺 | 241

We'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—is: 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 to or stronger than ours, and they dominate key industries.

But 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—

Eric, do you remember the first time we met?

Yes. Larry Page introduced me to you because he was on your board.

Yeah, so I got a call from Eric out of the blue, which was a great honor, and he 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.

I remember something—I love the story. We sat down to lunch at Charlie's Cafe. Of course, I'm running a nonprofit, and my mission is always raising capital for the nonprofit. So we sit down to lunch, and before we get started, you said, “Peter, what is your highest level of giving or membership?”

I said, “Well, Eric, it's our Vision Circle, for $2.5 million,” and you said, “Okay, I'm in. Now let's have a conversation.”

I'll buy them all. It was crazy.

The reason I wanted to come here is that this has become the epicenter of the abundance movement. And the abundance movement is correct. That's the important thing.

Thank you.

I want to thank you for all the support you've given me and XPRIZE all these years. I'm so grateful for that. I'll start with a question that I'd love to hear you expand on: We're living through a historic moment right now. Could you define the moment that we're in and give us 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. Some of it will happen, and some of it will take longer. For example, hardware takes longer than software, so robots take longer than digital systems on traditional hardware, and things like that.

The next thing that's really interesting—and terrifying, also—is recursive self-improvement.

Mhm.

It's not happening yet. 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 2. 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. 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, any single individual, or any corporation—we would still have advanced humanity because of these reasoning agents.

How fast do you imagine this is going to accelerate?

There's a thing which I call the San Francisco consensus, and the reason I call it that is because everyone in San Francisco believes this—everyone I know, anyway. 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 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 3 or 4 of these in my career.

In this thinking, once you have recursive self-improvement, where the system can begin to improve itself, you have intelligence learning on its own. In this argument, it will learn faster than we can because we're biologically limited.

The way this is expressed in San Francisco—and I'll give a simple example—is that you have a tech company with 1,000 fantastic AI researchers. One day, they turn on AI research—that is, an AI research agent. Well, how many AI research agents do you have? 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. 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 1 million of these agents. In AI, the way you determine you've made progress is that you have clear metrics showing that the reasoning, testing, or whatever the evaluation framework is, is better.

So that's what happens. In that scenario, the slope goes like this: You're already at this slope, then you add more people, then you get the agents, and you go like this. This is essentially a superintelligence moment.

The belief in San Francisco is that this occurs within 2 to 3 years. The evidence in favor goes something like this: Claude Code came out a couple of months ago—the latest one, Opus, whatever it is. What was it?

1.6, yes. Thank you.

Everyone I know in the Bay Area who's doing software says it was 80/20; now it's 20/80.

Mhm.

The best analysis I can come up with is that it's not the Claude Code part. It's that the underlying LLM can produce more reasoning over time and better-quality tokens over time. It's a deeper thinker, right?

Mhm.

All the labs are competing for that now. This is not just the size of the context window; it's actually the reasoning skill and the length of time for 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. I was a pretty good programmer. I watch what it does and I go, “My God, I'm over.” There's not a thing that I could do that it cannot do.

When they wrote a C compiler in Rust, I thought, “It's over.” So I think part of this is because the people who are building it are also seeing the diminution of their own skill. They'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.

Right?

The most likely scenario, by the way, has a lot of implications. 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 activities of this.

It also means that you're going to have a relatively small number of very large companies.

Yeah. Yeah. And this is a big deal.

Yeah. Yeah. And a very large number of very small companies, because you don't need as many people. You're watching that play out this month. This all happened in the last 3—

I was in one startup I'm involved with, and I was talking to the programmer, who was a perfectly brilliant young man. I said, “What's the truth?” He said, “Well, here's what I do.” He's working on UIs of various kinds, and I 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.”

I said, “What time?” He goes, “7:00 in the evening.” And I go, “Okay, what do you then do?” He has dinner with his wife, and he goes to sleep. I said, “Do you wake up?” He said, “No, I sleep very well.”

I said, “When does it finish?” “Oh, 4:00 in the morning.” Then he gets up, has breakfast, does whatever he does, and sees what's been invented. I mean, it's mind-boggling.

The stupid example I used with this young man shows the power of these systems: If you can define the evaluation function, you can let it run, and if you have enough hardware, you're inventing worlds. This stuff would have taken me 6 months and 10 programmers at Google to do the same thing, and this poor guy's sleeping.

It's so funny you say that, because I was literally backstage. They said, “Eric Schmidt's coming,” and I had my lid open on my Mac. 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.

I've got these 6 concurrent Claude 4.6 jobs open. You know, it's important what you're doing. Tony, interrupt for me. You're important, too, you know. It's crazy, because when we got together in Davos just 2 months ago, it was in this kind of autocomplete mode. You'd write the code, and then it would help you get it done. You were about 10 times more efficient, but you were still babysitting it.

Now, literally, it's working right now. When I get offstage, it will have solved 6 problems that I launched.

And I appreciate the excitement in the industry, but I can tell you, when I used to work on BSD—I basically worked at Berkeley on Unix, at Bell Labs, and on BSD Unix—we programmers invented what we needed. 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.” So one key thing to understand about digital intelligence is that the first inventors are the people solving their own problems—programmers. You shouldn’t be surprised by this; you should have expected it.

The other thing that’s interesting about programming is that it’s both scale-free, which means there are no particular limitations except electricity. You don’t need a lot of data; you already have GitHub and the equivalents. It’s also a fairly limited language set, so the number of language components, if you will, compared to human language is smaller. Smaller language, clear objective function—all you need is electricity.

Now, 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—quaint skills that we all used to have.

No, but 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 freshman men and women starting in September, which is a prompt-engineering class.

Why 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 on it. 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—they’re going, “This platform will be the expression platform for their art, their music, their writing, and so forth.” Why wouldn’t you teach them immediately? Peter, improve my proposal.

No, I just feel like AI is going to impact every student in high school today, and that they’re living an unnatural life by not engaging with it.

Well, plus your kids are that age, so they’re literally right now doing exactly what you’re describing. People here who have teenagers 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 restrictions. You really have to think about vulnerable teenagers with this technology. I did some analysis of where the real problems are with this stuff. A simple summary is that at some point there will be job impacts from this stuff. We’re seeing it in software and 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 right now.

For sure.

There are all sorts of other issues. The other one I came up with was 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.

Yeah.

So these are problems to be solved. We herald the future, and we solve the problems that it brought.

We’re going to talk about China, government, and jobs. But before we do that, I want to say I’m in this savor-the-moment kind of mode right now, because I feel like the world a year from today will be nothing like the world today. 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 1 minute about the fact that while you were running Google, the Transformer was invented there. The TPU was invented there. Demis Hassabis solved protein folding, which is now universally used. It does the work in an hour that used to take a PhD student 4 years.

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. You were there at the creation of everything we’re experiencing right now. Do you think anything like that profoundly strikes you about that moment?

Did you even realize at the time?

Still, 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 and pushed for excellence.

I’ll give you an example. In the early years of Google, my favorite interaction was one day when I said, “We need to hire some people doing Java.” Larry and Sergey said, “This is the stupidest idea we have ever heard.” I could never tell with them whether they were being serious or whether they were just joking with me. But their argument was that real programmers were programming one level lower. Today Google has many thousands of them.

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’d say, “Oh, that’s boring. Don’t do that. That’s another one of your ideas, right? We want a new idea.” I give them a lot of credit for it.

But what about the TPU in particular? I didn’t even hear about it until much later, and it takes years to design and build your own internal chips. Now it’s about to explode. I don’t know how much is public, but it’s just—

The 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.

For everybody’s benefit, inference is what the reasoning tech stacks I’m describing run on. So Google is particularly well positioned. As you know, NVIDIA purchased Groq for the reason of getting that inference capability.

Yeah, trying to catch up to what you thought of 10 years ago.

What’s interesting about NVIDIA, if you look at them—I was looking at the Rubin architecture—they 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 they will really be delivered to you. They just do the whole thing.

These are major industrial achievements, and that’s why both companies will do incredibly well.

Eric, in the AI exponential growth right now, talk to me about where the constraints are. You were in Congress talking about energy, chips, people, and capital. Where are the constraints right now?

It’s interesting. I started a data-center company with my friends. In my testimony, I said there was an estimated 92-gigawatt shortage of power in America between now and 2030. By reference, a nuclear power plant is about 1.5 gigawatts, so it’s about 60 nuclear plants, and we’re doing essentially 0 or 1, depending on how you count.

I got interested in the question of what the real resource constraint is in America, and it’s electricity. We have the universities. We have the smart people. We have the economics. We also have these amazing finance people who will give all of us billions and billions of dollars on a wing and a prayer.

There’s no country where the finance people are sufficiently crazy to do that. It’s not true in China. It’s certainly not true in Europe. These guys are incredibly jealous of the American financial system. So I always start by saying, “Thank you to the finance people for funding our dreams, whether they work or not.” Thank you.

There’s usually a retort at this point where people say, “Well, the algorithms will ultimately require less energy.”

I’m sure that’s true. There’s 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 6 to 9 months, it’ll be a bubble,” and so forth and so on. There’s no sign of this.

A team and I have been working on this for years. The ultimate scaling laws are not done yet. I keep asking my friends, “When does the asymptote arrive, and when does the curve slow down?” We have 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. That’s the great thing, frankly, about America.

Do you think it’s a limit to the capital, or a limit to where, 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, and data centers.

It's on the order of, depending on what numbers you use. So, 100 gigawatts—do the math.

Yeah. Can we raise $5 trillion over 5 years?

Yeah. That's the strength of America. Could we double that? The data center build-out is 1% of America's GDP growth.

We're back to a power problem.

Right. Well, thank you. The current estimate of electricity use in America is that 10% of the electricity in the United States will be used in 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. The standard data center that's being built is on the order of 400 megawatts. These things are, plus or minus, about half a mile long and about 500 feet wide.

They're essentially airflow machines. They 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 cool. Using NVIDIA as an example, the chips are water-cooled, and the HBM3E and now HBM4 memory put out so much heat that they have to be water-cooled. The chips are 2 kilowatts. I mean, this is insane. These things will kill you.

You want to hear something truly astounding and funny in hindsight? When you bought DeepMind, everybody thought it was like $800 million 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 today.

The credit for that one actually goes to Larry Page. Larry had studied AI when he was a Stanford graduate student, and we always deferred to him on this. He said, “This is the best team.” I think Elon and Larry competed over it. There was some complicated kerfuffle there.

Jeff Dean, who's the chief scientist, went over, and then he and I basically finished the deal. I still remember it: there, on one floor, were these sort of British people, led by a sort of Greek-British person, Demis. They were smart, but Google is full of other smart people. In 2016, Demis announced that we were going to win the game of Go. I figured, well—and by the way, at this point they were a separate group. We’d let them alone because they had to grow and figure out what they were doing and all that. This is the patience of capital. We could let them do that. We didn’t require that they do anything. So he said, “I’m going to go,” and I said, “Well, I’m going to come, too.” I flew to Korea, and it’s all one floor, and I met the team that had been winning the game. Of course, all of these Koreans were very excited about this because they knew they were going to beat the computer.

Mhm.

The Koreans were in one room, and I was in another. I went to the Korean room, and they were all saying, “We're going to beat the crap out of this Google group.” Then I went into the Google room, and it was very quiet. There was a monitor with what I now understand was an RL prediction mechanism showing whether we were winning or not.

It started at 50/50. I watched the Koreans talk for a while, and then I went to watch the screen. It went to 51%, and then it went to 52%. David, who was the architect, said, “Well, we just planned for it to get to infinity.”

[laughter]

Okay. So basically, it's the abundance theory. It's just—

[laughter]

And all the humans were crushed.

Yeah, they were all crying.

The DeepMind people said, “Yeah, yeah. They were supposed to.” Yeah, yeah. Welcome.

Then I understood the genius of the DeepMind people. 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 when we humans used to look for challenges where the computer could try to catch up, like chess. I think Go was the endpoint.

And we knew that, by the way. We understood that the game of Go was sort of incomputable by normal algorithms.

Yeah. There was lots of math that said you couldn't solve it. They came up with a two-tree model with 2 different RL trees.

One of the other things I learned about these systems is that it's not just, “Write me a Go program that will win the game.” You actually have to understand the game and so forth. They took the same team, and they got bored with Go after winning. So then they took the same team and had them work on protein folding.

Yeah. In protein folding, they took a whole bunch of protein scientists, which I know. Can we just think about the genius of that? Nobody would ever connect the game of Go to solving all of biology.

Demis was solving all of biology. But Demis had always wanted to work on this. Larry and Sergey were very interested in it, and protein folding is the perfect problem because you got a constant endpoint.

You got a defined endpoint.

So what happens is that people get excited about AI, but you need to have a validation function because these things don't have common sense yet. You have to show them what it's got. Ultimately, they produced this thing called AlphaZero, which essentially self-learns.

Let's talk about data centers in space. I'm in favor of it.

8 or 9 months ago, no one was discussing this.

I mean, all of a sudden. Do you know why I'm in favor of them?

I do, but you can mention it as you wish.

But all of a sudden, everybody's talking about them. What are your thoughts?

I'm part owner of a rocket company, and we need—

Which I love. I love having you come into the space community.

You understand this far, far better than I do.

Rocket science is named that for a reason. Rocket science is really, really hard. I don't know that much about rockets, although I certainly know how to manage tech people. But I think the opportunity is large and interesting.

There are challenges. There's an issue of getting heat off of it, because you don't have oxygen, and you also have radiation issues. Those have to get addressed.

But it makes the business plan for every rocket company that's large enough, right? The small guys aren't going to launch it, but Relativity Space, Blue Origin, and SpaceX—I mean, Elon's predictions were, I think, a launch per hour to populate the constellation he wanted. Do you think technically we're going to get there?

Well, the technology is understood. What's interesting about the data centers in space, technically, is heat dissipation.

Yeah.

That technology is understood. To me, it's a business question. Where should the data center be? Should it be in space, with these other issues, but with other benefits, including infinite power and so forth, versus on the ground, where you have fiber and it's not shaking too much?

Yeah. I mean, the energy argument says space wins by far, and I think the cooling is a very big challenge, but I think it's largely figured out now.

But then there's the politics of space. One of the turning points in AI history was you getting in front of Congress and saying, “Hey, we need to find almost 100 gigawatts.” At the time, it seemed outrageous, and now, of course, it's mainstream. It looks like the crazy investors are solving the problem. The unleashing of the money is happening.

American capitalism. The tech industry—I mean, we have another set of problems.

Well, if the next frontier is space, then there's no investment community in space, but there's also no military jurisdiction in space—or maybe there is. I would never have thought my childhood dreams of going to the Moon and Mars would be fueled by data centers.

Mine can't carry you at the moment, but it can in the future, maybe.

Yeah. Yeah, all the way. I read a Time op-ed piece you wrote last night: China can dominate the physical AI future—or what did you say? Can you summarize that for us? It was an important conversation.

The geopolitical context. I've said this many times, and I'll say it again: The American competitor—not enemy, but competitor—is China.

I think it's—by the way, I think it's an important distinction for you to make, so thank you.

Not enemy, competitor. How do we understand them as a competitor? They have lots of money. They're very, very smart. Their work ethic is equal to or stronger than ours, and they dominate key industries.

With respect to robotics, we somehow decided it was okay for them to dominate the electric vehicle industry. This was an error. To be very clear, it's an error. Why do you not understand it's an error? Because we don't allow their cars in?

Spend some time outside of this country in Chinese cars, trust me. They are real competitors. They've done a great job.

As I understand it, China is capable of vertical integration and building these gigafactories at a scale that we can't, for all sorts of reasons. That's got to get addressed. So, if you want to compete—and I want to compete and win with China—I want us to have the same kind of system. I want to compete, not be an enemy.

In robotics, it turns out you can understand robots as essentially actuators: these little stepper motors—click, click, click—and a brain. Ignoring the appearance and the googly eyes and all that kind of stuff, the electric-vehicle industry produces the same kind of motors and the same kind of systems. They have an expertise that we don't.

My own view is that, at least at very low cost, China is going to win that. That was what I was trying to say in that piece, and I worry about that. Today, these are not particularly useful. They're fun toys—a replacement for the dog if you get mad at your dog. Sorry, I love dogs, but you get the idea.

We need to address this. At the moment, it sure looks to me like China's robotic hardware is the winner at the low end. I'm not talking about the high end, the expensive stuff, or industrial robots. If you're confused, watch the Unitree robot dance with the humans. That came out about a month ago.

Unitree is here in the tech hub, and the co-founder will be on stage with us later today. Pay attention to them. They're very impressive, and I spent some time with them the last time I was in China. They're one of many.

The way China works is that they have brutal competition—brutal. It's unbelievable. I was talking to my friend—we teach at Stanford—and he said, “In China, we don't have the board dinner. We have a 2-hour meeting. We get back to work.” There's no preamble. We're not saying hello, asking how you are, or asking about the family. We're boom, boom. It's just cultural.

The work ethic, the precision, and the scale that is possible in China are a real competitive advantage. 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.

Interesting. The Chinese model very much has a well-built-out supply chain, with many vendors in the loop.

But we've been on a worldwide tour of all the humanoid-robotics companies. By coincidence, I guess, when you look at the gigafactory and look at Elon's vertical integration, and also at Brad Adcock, it's the same thing: it's all vertically integrated. Why?

Well, because there's no vendor. I have no choice.

So, it appears that to get to abundance—again, we're in the Abundance Group, this is the Abundance Club—the way you get to abundance is that you drive prices down and get vertically integrated.

Elon, in our country, pioneered that, to his credit.

Yeah. The old joke about Google was, “We would build anything, including the buildings.” Well, Elon is actually doing that.

Yeah.

Why? Not because he's insane, but because that's how you drive costs down. He truly believes—and I'd love to get your take on this—that the robot building the robot is imminent. I didn't get it until we toured the gigafactory and I realized that almost all of it is automated already. The last piece is the human controlling a few knobs, and the humanoid robot can actually do that.

Let me define the boundary, because it's important. Let's use batteries, for example. Batteries are predictable, straightforward manufacturing processes at huge scale. Things like that will have gigafactories.

One of the questions—and I, of course, used LLMs to do my deep research as a new person in the rocket company—was how much of the human labor to build a rocket could be replaced by robots. The current limit, which of course will change, is that in our company we have these extraordinarily talented assembly people. They're more than welders and more than mechanics. They understand precisely how the tubes and so forth go together, and they use precise tolerances. That kind of assembly is beyond current robots. I'm sure it will eventually show up, but not for a long time.

People don't realize the majority cost of a rocket is labor.

Exactly. When they get inside the rocket, they understand what they're doing, they see what's wrong, and they use human judgment. We don't have those systems yet. Perhaps we will in the future, or perhaps this will be one of the last things to go. At the moment, high-skilled mechanical labor is very important. Low-skilled labor of any kind gets swept up.

So, both of those involve self-improvement: AI self-improvement, and then robotics—building robots that build robots—self-improvement. Both of those are loops. We talk a lot about closed loops being—

If I can interrupt you.

Yeah, please.

The term I like to use is “learning loops.” In a business, try to figure out all the different learning loops and then try to accelerate the learning. Fastest learner wins. Sorry.

I'm going to lead the witness here a little bit. But if you asked Daniela Rus over at CSAIL or Erik Brynjolfsson at Stanford a year ago, “Do we need one more big breakthrough in core AI science, or will scaling what we've already got lead to self-improvement, and then will that be all we need to get to AGI?”

And then self-improvement. I spent the last week doing RSI reviews—recursive self-improvement reviews. The scientists do not agree on the exact approach that will work yet, so I think it's too early to know the answer to that question. There's evidence that it will work. There are tests in the lab that show it, but they show it in limited cases that are kind of demos.

Real recursive self-improvement is the following: start now, learn everything, discover things, and tell me what you learned. That query doesn't work yet.

We're seeing all of the frontier labs constantly leapfrog each other. Literally every week, it's a new model. Unbelievable. It's extraordinary. Do you imagine they're all converging toward the same endpoint, or is anyone going to pull ahead?

If you go back to this question of capital, how much room is there in the world for these companies? How many can there be? I'm going to make some numbers up, and these are made up. I think there's at least 10 in the world at this scale.

I think there will be a few in China. I think the majority will be in the United States—the usual suspects, most likely. There might be 1 or 2 in Europe, depending on their electricity costs, which are a problem. There might be 1 in India. There's not going to be 1 in Russia because of the war, and so forth.

Can the world accommodate 10? Yes. Do they track together? I don't think so.

One of the key things to understand about China is its approach, which has produced DCV4, Quinn, Kimmy 2, Kimmy version 3, and more coming. It's all open source and open weights. They've managed to do this with our chip limitations against them, which annoys them no end. It shows you how clever they are.

The Chinese strategy is also a bit different. It's less centralized computing and much more edge computing, which has to do with enveloping their Chinese customers with AI around them all the time. We're much more AGI- and ASI-centered, which is fine. The patterns are diverging.

Within the companies, it's a jumble now. Fundamentally, Microsoft and Google have these large cash-flow streams from enterprise, so they can fund that. Anthropic has done a fantastic job of raising money from those other companies, too, and they use, as you know, Google TPUs. They've become the leading player in the Claude API within the enterprise-agent system.

OpenAI is now shifting some of its strategy to include the new things it's doing. I don't think we can predict. The key thing to understand is that they need so much money. You look at what Sam is trying to raise—they need so much money that they're forced into these situations where they have to win those battles. They're busy winning them. This is all good.

In a year, we'll know better the answer to your question.

I have 2 questions I want to close this out with that I think are important. You made a statement a couple of times—once in our podcast, once elsewhere—that regarding AI safety, the world may need to have a modest Chernobyl-like death event in order for us to wake up. Do you still believe that's the case?

Yes. By the way, I'm not endorsing that. I'm describing it, not prescribing it. What are the real dangers of this? There are biological dangers. There are obviously dangers to kids and democracies and so forth.

But let's think about a biological attack or a nuclear attack that's spawned by these things. It may take such a tragedy—hopefully a small one—to awaken the world and help us understand that these things have negative power.

I can imagine—I'm making this up—that something bad happens, and then all of the leaders—China, the United States, everyone—have a meeting and basically say, “What are we going to do?” We're in brutal competition, we hate each other, I don't like you, you don't speak the same language, and so forth. But we are all in it together over this issue.

My sense is that will happen, but I don't know when. We had a congressperson on this stage last year, and someone in the crowd asked, “How much time do you spend talking about AI in Congress?” He said, “Well, it's definitely way less than 1%.”

Yeah. And so, without that wake-up call, I don't see how you have—

Governments are super busy, right? They're driven by political things and, in democracies, by political sentiment. I'd like us as a nation to focus on the following: I want to win the AI race. I want us to do whatever it takes to do that.

This government is doing a very good job of making energy permitting more accessible. The rate at which data centers are getting built has now accelerated, solving the grid problems. I also want lots of immigrants in our country because those immigrants—at least the high-skill immigrants—are what we need. We need the smartest people in the world on our side to build these systems. This is a unique moment in history, right?

I would take us home on a positive note. You said we're going to get to ASI at some point, whether it's 2 years out, 5 years out, or somewhere in this next decade. The question is, what steps can we take to steer artificial superintelligence toward abundance, toward uplifting humanity, and in alignment with humanity?

To make this abundance thesis materialize, what's your advice to us—companies and governments? There's an over-reliance in our society on people like me to work on this. Why don't we have the smartest people in politics, history, human psychology, governance, and ethics working together to make sure this stuff stays aligned with human values and human alignment?

I want the system that we build in America to reflect American values: the values of freedom, freedom of speech, and freedom of association—all those things you learned in elementary school and high school. They're still important to our nation. They're the enablers for the next generation of our children and grandchildren. I desperately want that, and I don't want America to ever get on the wrong side of that battle.

There's lots of people working on this. Lots of people understand the technical details. I happen to run an informal group that discusses this every week. So it's possible, but it requires political will and an understanding that this can be done without screwing up the genius of America, right?

In other words, I'm not suggesting slowing anything down. I'm—you said it so well—shaping it. Making sure we don't cross lines. Like I already mentioned, the underage kids problem. That's a line we can't cross. We have to solve that problem. There are others.