第1部分：Eric Schmidt 与 Fei-Fei Li：人工超级智能之后的人类生活｜EP #206

What does superintelligence mean, and what happens when it arrives? We've been talking about AI, AGI, and now perhaps digital superintelligence, or ASI. I want to start with the obvious question, and it's one that I don't think anybody has a perfect answer for: What does superintelligence mean, and when is it likely to be here? Eric, we've talked about this. What are your thoughts?

Thank you, Peter, and thanks to everybody for being here. Obviously, thanks to Fei-Fei Li, our very close colleague.

The generally accepted definition of general intelligence is human-level intelligence, or AGI. Human intelligence you can understand because we're all human: You have ideas, you have friends, you think about things, and you're creative. Superintelligence is defined as intelligence equal to the sum of everyone, or even better than all humans.

There's a belief in our industry that we will get to superintelligence. We don't know exactly how long. There's a group of people whom I call the San Francisco consensus because they're all living in San Francisco. Maybe it's the weather or the drugs or something, but they all basically think that it's within 3 to 4 years. I personally think it'll be longer than that. But fundamentally, their argument is that there are compounding effects that we're seeing now which will race us to this much faster than people think.

And Fei, I don't think anybody expected the performance that AI has given us so far. The scaling laws have given us capabilities that are extraordinary. You're the CEO and founder of a new company, World Labs, and you've been at Stanford working on this. How do you think about superintelligence? Do you discuss superintelligence at all in your work?

That's a great question, Peter. When Alan Turing challenged humanity with the question, “Can we create thinking machines?” he was thinking about the fundamental question of intelligence. The birth of AI is about intelligence and the profound, general abilities that intelligence entails. From that point of view, AI was born as a field that tries to push the boundary of what intelligence means.

Fast-forward 75 years after Alan Turing, and this phrase, “superintelligence,” is pretty hot in Silicon Valley. I do agree with Eric that the colloquial definition is the capability of AI and computers that's better than any human.

But I do think we need to be a little careful. First of all, some parts of today's AI are already better than any human. For example, AI's ability to speak many different languages and translate between dozens and dozens of languages—pretty much no human can do that. Or AI's ability to calculate things really fast, and AI's ability to know about everything from chemistry to biology to sports—the vast amount of knowledge. So, it's already superhuman in many ways.

But it remains a question: Can AI ever be Newton? Can AI ever be Einstein? Can AI ever be Picasso? I actually don't know. For example, we have all of the celestial data of the movement of the stars that we observe today. Give that data to any AI algorithm, and it will not be able to deduce Newtonian laws of motion. That's an ability that humans have. It's the combination of creativity and abstraction. I do not see today's AI, or tomorrow's AI, being able to do that yet.

One of the common examples—and Fei, of course, got it right—is to think about whether, if you had all of the knowledge that existed in 1902 in a computer, you could invent relativity, basically the physics of today. The answer today is no.

For example, if you look at what is called test-time compute, where the systems are doing reasoning, they can't take the reasoning that they learned and feed it back into themselves very quickly. Whereas if you're a mathematician, you prove something, and you can base your next proof on that. That's hard for the systems today, although there are approximations.

We don't know where the boundaries are. The example that I'd like to use is this: Let's imagine that we can get computers that can solve everything that we normally can do as humans, except for these amazing sets of creative abilities. How do really creative people do it? The best examples are that they are experts in one area, they see another area, and they have an intuition that the same mechanism will solve a problem in a completely different area. That's an example of something we have to learn how to do with AI.

An alternative would be to simply do it by brute force using reinforcement learning. The problem is that, combinatorially, the cost of that is insane, and we're already running out of electricity and so forth. So I think that to get to real superintelligence, we probably need another algorithmic breakthrough.

We need another what?

Algorithmic breakthrough—another way of dealing with this. The technical term is called non-stationarity of objectives. What's happening is that the systems are trained against objectives. But to do this kind of creativity that Fei is talking about, you need to be able to change the objectives as you're doing them.

We've seen this past year, I think GPT-5 Pro reached an IQ of around 148, which is extraordinary. Of course, there is no ceiling on this. It loses meaning at some point, but the ability for every human on the planet to have Einstein-level intelligence—not on the creativity side, but on the intelligence side—in their pocket changes the game for 8 billion humans.

Now, with Starlink and $50 smartphones, it's possible that every single person on the planet has this kind of capability. Add to that humanoid robots. Add to that a whole slew of other exponential technologies. The commentary is that we're heading toward a post-scarcity society, right? Do you believe in that vision, Fei?

I do think we have to be a little careful. I know that we're combining some of the hottest words from Silicon Valley: AI, superintelligence, humanoid robots, and all that. To be honest, I think robotics has a long way to go. I think we have to be a little bit careful with the projection of robotics.

The ability and dexterity of human-level manipulation—we have to wait a lot longer to get it. So, are we entering post-scarcity? I don't know. I'm actually not as bullish as a typical Silicon Valley person because I absolutely believe AI will be augmenting human capabilities in incredibly profound ways. But I think we will continue to see that the collaboration between humans and AI will be the most productive and fruitful way of doing things.

The projection is that AI is going to generate as much as 15 trillion in economic value by 2030. The idea is that it's shifting the foundation of national wealth from capital to labor to computational intelligence. What's that implication, Eric, for the global economy? How are we going to see redistribution, if you would, of wealth or of capabilities? Are we going to see a leveling of the field between nation-states, or are we going to see runaway winners?

In your abundance hypothesis, which we've talked a lot about, there may be a flaw in the argument because part of the abundance argument is that it's abundance for everyone. But there's plenty of evidence that these technologies have network effects, which concentrate benefits among a small number of winners. You could, for example, imagine a small number of countries getting all those benefits within those countries. You could imagine a small number of firms and people getting those benefits. Those are public policy questions.

There's no question the wealth will be created because the wealth comes from efficiency. Every company that has implemented AI has seen huge gains. Think about where we are here in Saudi Arabia: You have all of this oil distribution, all the oil networks, all the losses. AI can easily improve that by 10% to 20%. Those are huge numbers for this country.

If you look at biology, medicine, and drug discovery, you have much faster drug-approval cycles and much lower-cost trials. Look at materials: much more efficient and easier-to-build materials. The companies that adopt AI quickly get a disproportionate return.

The question is: Are those gains uniform, which would be our hope, or, in my view, more likely largely centered around early adopters, network effects, well-run countries, and perhaps capital?

But you could imagine that we're going to see autonomous cars in which being in an autonomous vehicle is 4 times cheaper than owning a car. We can see AI giving us the best physicians and the best healthcare for free, in the same way that Google gave us access to information for free. We will see a massive demonetization in so much of our world.

I think that will be available to anyone with a smartphone and decent bandwidth connectivity. Is that still not what you think will happen? Do you think there's a reason something would stop that level of distribution of those services, which we spend a lot of our money on today?

I do think AI democratizes that. I totally agree with you. I think whether it's healthcare, transportation, or knowledge, AI will democratize massively. But I agree with Eric that this increased global productivity does not necessarily translate to shared prosperity.

Shared prosperity is a deeper social problem. It involves policy. It involves geopolitics. It involves distribution, and that's a different problem from the capability of the technology.

So, what's your advice to the country leaders who are here, who are seeing ASI as a future for someone else and not for themselves? What should they be doing? I mean, this is critical—the speed at which it's deploying. They don't have a lot of time to make critical decisions.

Well, it's worth describing where we are now in the United States. Because of the depth of our capital markets and because of the extraordinary chips that are available from Taiwanese manufacturers—TSMC in particular—America has this huge lead in building what are called hyperscalers. If there's going to be superintelligence, it's going to come from those efforts. That's a big deal.

If there is superintelligence, imagine a company like Google inventing this, for example. I am obviously biased. What's the value of being able to solve every problem that humans can't solve? It's infinite.

Sure.

So, that's the goal, right? China is second. It doesn't have the capital markets, it doesn't have the chips, and the other countries are not anywhere near. Saudi has done a good job of partnering with America, and the hyperscalers will be located here and in the UAE. That's a good strategy.

That's a good example of how you partner. You figure out which side you're on—hopefully it's the United States—and you work with the U.S. firms. I do think countries all should invest in their own human capital, invest in partnerships, and invest in their own technological stack as well as the business ecosystem.

As Eric said, this depends on the strength and particularity of the different countries, but I think not investing in AI would be macroscopically the wrong thing to do.

So, under the thesis that investment involves building out data centers in your nation, do you think every country should be building out a data center with sovereign AI running on it?

“Every country” is a very sweeping statement. I do think it depends. It depends. I think, obviously, for a region like this, absolutely, where energy is cheaper and it's such an important region in the world. But if we're talking about smaller countries, I don't know if every single country can afford to build data centers. But there are other areas of investment, right?

But let me give you an example. Let's pick Europe. It's easy to pick on Europe. Energy costs are high, right? Financing costs are not low. So, the odds of Europe being able to build very large data centers is extremely low. But they can partner with countries where they can do it.

France, for example, did a partnership with Abu Dhabi. So, there are examples of that. If you take a global view and figure out who your partners are, you have a better chance.

The one that I worry a lot about is Africa. The reason is: How does Africa benefit from this? There's obviously some benefit of globalization—better crop yields and so forth. But without stable governments, strong universities, and major industrial structures, which Africa, with some exceptions, lacks, it's going to lag. It's been lagging for years. How do we get ahead of that? I don't think that problem is solved.

We've seen incredible progress with AI today, effectively beginning what people call solving math. That potentially tips physics, chemistry, and biology. We have the potential—my time frame is the next 5 years; others may think longer—to be in a position to solve everything, where the level of discovery and the level of new product creation, new materials, biological therapeutics, and such begins to grow at a super-exponential rate.

How do you think about that world in 5 years, Eric?

So, first, I think it's likely to occur, and the reason technically is that all of the large language models are essentially doing next-word prediction. If you have a limited vocabulary—which math is, and software is, and cyberattacks are, I'm sorry to say—you can make progress because they're scale-free. All you have to do is just do more.

If you do software, you can verify it. You can do more software. If you do math, you can verify it, and do more math. You're not constrained by reality, physics, and biology. So, it's likely in the next few years that in math and software, you'll see the greatest gains.

We all understand your point that math is at the basis of everything else. I think Fei-Fei is the expert on the real world. There's probably a longer period of time to get the real world right, which is why she founded the company of which I'm an investor. Do you want to talk about that?

Yeah. Well, first of all, I actually want to respectfully disagree. I do not think that we will solve all the fundamental math, physics, and chemistry problems in 5 years.

We're going to take a bet on that one.

Yes. So, 50-50?

Okay, you got it.

We should take a bet on that. Part of humanity's greatest capability is to actually come up with new problems. As Albert Einstein said, most of science is asking the right question. We will continue to find new questions to ask, and there are so many fundamental questions in science and math that we haven't answered.

Fei-Fei, your new company, World Labs, is creating extraordinary, persistent, photorealistic worlds. Are you expecting that we're going to be spending a lot more of our time in virtual worlds? My 14-year-old boys right now are spending way too much time in their virtual gaming worlds.

But is this what we're going to do in 10 or 20 years, in a post-ASI world where we don't have to work as much, we have a lot more free time, and our robots maybe by then are serving us? Are we going to live in virtual worlds?

Great question. What we are doing is building large world models. That's the problem after large language models: Humans have the ability to have the kind of spatial intelligence with which we can understand the physical 3D world, imagine any kind of 3D worlds, and be able to reason and interact with them.

Up until what our company has been doing, we did not have such a world model. World Labs, the company I co-founded and am CEO of, has just created the first large world model.

The future I see—I actually agree with you—is that we will be spending more time in the multiverse.

Yes.

Of the virtual worlds. It doesn't mean that reality—the real world, this world, this physical world—is gone. It's just that so much of our productivity, our entertainment, our communication, and our education are going to be a hybrid of virtual and physical worlds.

Think about medicine: How we conduct surgery is very much going to be a hybrid world of augmented reality, virtual reality, as well as physical reality. We can do that in every single sector. Humanity, using these large world models, is going to enter the infinite universe.

I had a chance to see your model backstage. It's amazing. The technology Fei-Fei is building is going to be world-changing.

So, my last question here is about human capital. Superintelligence has been called the last invention humanity will ever make, as it could eventually automate every process. We'll see if it automates discovery. We'll see how much of creation it automates.

But in a world where the best strategic, scientific, and economic decisions are being made by machines at some point, what is the ultimate irreplaceable function of human intellect and leadership? What are humans innately going to be left with in 10 or 20 years?

Well, in 20 years, we will enjoy watching each other compete in human sports, knowing that the robots can beat us 100% of the time.

But if you go to Formula 1, you're going to want to see a human driver, not an automated car.

Yes. Humans will always be interested in what other humans can do, and we'll have our own contests. Perhaps the supercomputers will have their own contests, too.

But your reasoning presumes many, many things. It presumes a breakout of intelligence in computers that's humanlike—unlikely, probably a different kind of intelligence. It presumes that humans are largely not involved in that process—highly unlikely.

All of the evidence—and Fei-Fei said this very well—is going to be human-computer interaction that basically we will all have. So, going back to what you said about 8 billion people with smartphones, with Einstein in their phone, the smart people, of which there's a lot, will use that to make themselves more productive.

The win will be teaming between a human and their judgment and a supercomputer and what it can think and remember. There is a limit to this craze: Supercomputers and superintelligence need energy.

So perhaps what will happen at some point is that the supercomputers will say, “Huh, we need more energy, and these humans are not building fusion fast enough.” So we’ll accelerate it. We’ll come up with a new form of energy. Now, this is science fiction, but you could imagine at some point the objective function of the system says, “What do I need? I need more chips or more energy, and I’ll design it myself.” Now, that would be a great moment to see.

I agree.

I do want to say it’s so important, as we talk about AGI and ASI, that the most important thing we keep in mind is human dignity and human agency. Our world, unless we are going to wipe out this species—which we’re not—has to be human-centered. Whether it’s automation or collaboration, it needs to put human agency, dignity, and human well-being at the center of all this. Whether it’s technology, business, product, policy, or any of that, I think we cannot lose our focus from that.