前 Google CEO 拆解中美 AI 竞赛，以及如何避免全球危机——Eric Schmidt 博士与 Dave Blundin | EP #207

Is America going to win the AI race?

There are 3 obvious threats right now.

Is China winning the global race to develop artificial intelligence technology?

So, how far ahead of China do you think we are?

China put in 172 gigawatts of solar last year, I think is the number. It's remarkable. We needed, in our calculation, by 2030, 92 gigawatts to be built. A big nuclear power plant is somewhere between 1 and 1.5 gigawatts.

Now that's a moonshot, ladies and gentlemen.

Everybody, welcome to part two of Eric Schmidt Week on Moonshots. In this episode, my moonshot mate, Dave Blunden, is interviewing Eric Schmidt about US versus China and how to avoid crisis during this period of hyperexponential growth in AI. Heads up, uh, this, this audio recording from Eric is a little bit choppy. He was on Wi-Fi from his hotel room. But guarantee you, the content is valuable, so please listen in. And also, this was recorded about a month ago. It took us a while to get the footage out to all of you. All right, let's jump into this episode with Dave Blunden and Eric Schmidt.

Is America going to win the AI race? I know that's a topic that you've spoken on quite a bit. It's also addressed a little bit in your new book, Genesis: Artificial Intelligence, Hope, and the Human Spirit, which hopefully everybody will read, with Henry Kissinger as a co-author. Are we going to win the AI race? What are the scenarios where we win or lose?

1. America Leads the AGI Race

It looks like we will. Let me define it. I think the San Francisco Consensus, as I call it, is what people in San Francisco believe, which, charitably, was you. You're going to see a build from current agentic computing to various forms of recursive self-improvement, to eventual AGI and superintelligence.

In order to do that, it requires an enormous amount of hardware—Google TPUs, big chips, and so on. Everybody in the audience knows that. It sure looks like the hardware restrictions that the Trump and Biden administrations have put on China are going to prevent them from competing in that space.

I've been recently in Shanghai for a few days. I have good relationships with the Chinese, and my conclusion is they're fighting a different game. They're going to build out AI in every product, every service, everything, but in a more classical way, whereas America is going to seek AGI.

I was quite worried that we would end up in a superintelligent race where you would end up with such enormous gains that one side would have to actually attack the other one—in other words, a preemptive attack. It looks like that fear of mine was not well grounded in facts. I think we're going to be okay for a few years.

Really? What about robotics? We seem to be pretty far behind on that front.

If it's okay for me to be completely blunt—

Please.

2. China Dominates Robot Hardware

The Chinese are doing the same thing in robotics that they have done in electric vehicles. In case you're confused, while the current government in the US gets rid of solar and wind subsidies and promotes oil and gas, China put in 172 gigawatts of solar last year, I think is the number. It's remarkable.

The Chinese race around solar and EVs—electric vehicles of one kind or another—and it looks like they've won. They're using all of those technologies, in particular, new ways of building stepper motors and other very inexpensive, very powerful physical things.

A good example is Unitree, which just launched the R1 for 6,000 dollars, available in December. I've ordered one. We'll see how good it is. But the arrival of humanoid robots is likely to be dominated by China.

I'm not suggesting there won't be areas where the US will play. The US will still have spaces of very high-end, very sophisticated stuff. Our software is so much better than the Chinese software. But at the hardware level, I think you should assume that the world will be awash in inexpensive Chinese robots—

Interesting.

—in the same sense that they'll be awash in inexpensive Chinese electric vehicles. Who knew?

This is going to be my last question, but let me bump it to the front since you just beautifully segued into it. In the race for AGI, but also in parallel robotics, China is going to be miles ahead in electricity production.

In the very short term, we're all going to be chip-constrained, TPU-constrained. But if you look 3 or 4 years out, the fabs are running at full throttle, the chips are coming out by the millions, and then you're suddenly electricity-constrained. Is there a vulnerability for America there?

3. Electricity Becomes the Bottleneck

It's a huge issue. Again, let's use China versus the US as a metaphor. What are China's strengths? I'm not praising China; I'm just trying to report it. They have solved their electric power problem. They also have full control over social media, so they don't have the kind of problems that people here complain about. They have enormously talented software people, and they don't have enough hardware. I think that's roughly where they are.

They're also incredible capitalists. They call it “socialism with Chinese characteristics,” but trust me, it's pure, raw capitalism. Let's call it what it is.

In the US, we have the many benefits everybody understands. We do not have enough electricity, and at least in our consumer stuff, the Chinese are likely to beat us. Our hardware architectures are fantastic, and I'm including the Amazon chip, obviously, the embedded chips, TPU v5, which I'm happy to say I was part of in TPU version 1. All of that stuff is incredible.

So, if you think about it—and I testified in Congress a month or 2 ago on this—we looked at the amount of electricity required in the United States to power the expected demand of data centers, and we needed, in our calculation, by 2030, 92 gigawatts to be built. For reference, a big nuclear power plant is somewhere between 1 and 1.5 gigawatts.

To give you a sense of how many nuclear power plants are getting started in America: effectively zero. So we had hoped—and I've hoped in my lobbying and testifying—that the government would fast-track the availability of all kinds of electricity.

Indeed, they have promoted oil and gas, but they've also hobbled solar and wind to a terrible degree, which is an error. The country needs more energy, and if we don't get more energy, we're not going to be able to fully exploit the lead we have in AI and AGI. It's very clear.

By the way, the obvious next question is, what do we do? There are scenarios. For example, the president went to Saudi Arabia and the UAE and did huge deals for multiple gigawatts. We might find ourselves in a situation where our training for our most important thing—the thing which is the essence of America, which is American intelligence—is actually being developed in kingdoms, and that may be the only fallback we have.

That is really weird, and that begs a very difficult question. Do you mind if I ask you a tough one?

Go ahead.

I've been inspired by you for most of my adult life. I've really been watching your podcast where you're talking about this. Do you remember when you guest-lectured Eric Brynjolfsson's class at Stanford? You said that when you were running Google, you felt like you made decisions 3 times faster than any company on the planet, but then when you got into the federal government, you felt like the decision-making was one-third as fast as even a slow company.

From your point of view, it's a tiny fraction of the pace that we need to move. But you've been saying recently that one of the things that's inevitable is some kind of an AI disaster, and we're hoping that it's 100 people who die and not 1,000, 10,000, 1 million, or even 100 million.

Suppose that it does play out, that there's a catastrophe that's a wake-up call. What do you want to do with that wake-up call? What's the next move for Eric Schmidt after the wake-up call?

4. AI Requires New Treaties

The background here is that Dr. Henry Kissinger and I spent an awful lot of time talking about the period in the 1950s when he was a key component of all of these things. What he did was use the fact that we had used the nuclear bomb to negotiate, over about a 15-year period, a set of treaties that restricted nuclear proliferation.

Those treaties, when they were negotiated, have allowed us to be alive today. They were centrally important. Without controlling the spread of enriched uranium and the other secrets, we would all be toast, literally, because of crazy people and so forth.

Is there an analogous set of things that we can do? The problem here is—or I guess the good news is—we're not in a war. We haven't had a nuclear bomb. We don't have that thing to discuss. So we can talk about it, but governments tend to act reactively.

There are 3 obvious threats right now, which I think are fairly well understood. The first is misinformation, and the software that we're all collectively giving people allows for all sorts of misinformation, fake videos, fake news, what have you.

We all understand this. It's all open source. That's done. That's a threat to democracies and maybe to dictatorships, but certainly to democracies.

The second one is cyber, and I think one way to understand cyber is that if you can write code, you can also write cyberattacks. It's the same logic, and you have these incredible gains in software. It's frightening how good these software systems are. Remember, my career is as a programmer. These things program better than I ever did. It's shocking.

The third one is bio. I think most people believe that one of those 3 will create some kind of mini-crisis that will then cause governments to say, “Hang on. Let’s have a conversation about how to really deal with the downsides.” The upsides are incredible. I want America to win, and I want us to run as fast as we can. We’re indeed doing that with the Trump administration, which is great. But we have to be aware that these things are possible.

The one I’m particularly worried about is biological, and it goes something like this: You take some existing pathogen, and using biological techniques, which I won’t discuss, you can modify it enough that it cannot be detected, but it’s still quite dangerous. That’s an example of a threat. There are many others.

Yeah. That’s also the easiest, I think, which is scary, of the immediate CBR and anthrax threats, so cyber, biological, radiological, and nuclear. Biological is the one you can do in a basement with 3 people, but it’s also the one that’s hardest to contain.

I hope the theory is right. How do you deal with proliferation? The Biden approach was, “Okay, let’s contain AGI to these 5 companies,” with Google being one of them. “And then let’s say any model with over 10^26 training FLOPs has to register with the federal government, and then we’ll keep it all contained.”

That all got scrapped immediately after the election and was replaced by the new David Sacks document. The David Sacks document is much more about how we move as quickly as possible and win the race, but it doesn’t really address proliferation. Obviously, in America, we want startups and researchers to have incredible access to technology and compute.

On the other hand, 3 people in a basement making a biological weapon is a really scary thing. How do you balance those?

It turns out, if you read the David Sacks–President Trump announcement, they’re very clear that they want to continue to study the security aspects of AI, especially in a geopolitical way, and that’s code for China versus the US. They continue in their proposal to fund things involving nation-state attacks and so on, and I fully support that kind of stuff.

There’s probably a difference on the misinformation and social media stuff between the 2. For example, the Biden rule was simply that you had to report if you were doing a training run of 10^26 or greater. I was part of the group that made that number up. We made it up because we had no better number. I’m not suggesting it’s the right number.

The conclusion you come to is that it’s probably the case that our government—I don’t know this, but I’m guessing—knows where the training runs are going on in China because of espionage. It’s probably the case that the Chinese have espionage on the US, knowing where our training runs are. So I’m not sure the nature of the training runs is a secret. Frankly, everybody knows where the data centers are because they’re immense.

Another attack on 10^26 is that the training is getting more cost-efficient. If you look at the move from something called FP16 to FP8, that means 8-bit floating point. People are now moving to 4-bit floating point, which is bizarre. It turns out these training algorithms seem to be quite tolerant of floating-point imprecision, which is a shock to me.

Again, we’re getting more efficient in training. This creates more of a proliferation problem. If I can state the proliferation issue in general, I’m not worried about big companies and big countries because I can count them. There’ll be 10 huge data centers and 10 huge training runs around the world.

I’m much more worried about the open-source groups, which can operate in the shadows. They don’t have to solve every problem; they just have to do one thing well. They can patch together the open source, which is generally available, and it’s good enough.

If you look at the quality of DeepSeek R1, and now R2 coming, it sure looks like it’s at 80% or 90% of these top closed models. The people behind the closed models, which obviously includes Google, get very defensive over this because they say, “Look, those are kind of synthetic. They’ve diffused the models. They’ve trained on our best information.” All of that is true. They’re correct, but they are nevertheless useful for specific things, and they could be used in proliferation scenarios to carry out various forms of cyber and biological attacks.

The current rule of thumb is that distillation and transfer learning are about 1% of the cost of the original training that gets you to the same destination. If you take FP4, which you mentioned, you get an 8× performance boost over FP32. So you get an 8× on the weights, and then you’ve got another 100× on the transfer learning.

It’s really hard to draw an analogy to nuclear or chemical weapons in the past because you couldn’t shrink them 1,000× under the covers and get the same result. But AI is really weird that way. It’s very compressible, very fluid.

Well, again, we’ll see. It does not look like we have a very good solution for distillation. It looks like an opponent of a company can mask their queries to look like normal sets of users and then distill the models. One of my friends has thought about this a lot, and he thinks that the eventual state in the United States is that the biggest models will never be released, and that the companies will distill their own models for that reason. That’s his opinion. We’ll see if that’s true.

Mm.

This produces a bizarre outcome where the biggest models in the United States are closed source, and the biggest models in China are open source. The geopolitical issue there, of course, is that open source is free and the closed-source models are not free.

The vast majority of governments and countries that don’t have the kind of money that the West does will end up standardizing on Chinese models, not because they’re better, but because they’re free. We’ll see if that’s true or not, but I do worry about that.

This whole topic of distillation and proliferation is a good segue into a much happier topic, which I really want to use our remaining time on. I cannot tell you how inspiring you are to the founders around Cambridge, MIT, Harvard, and Northeastern, where all of our talent comes from.

In my perfect world, you would podcast or say something or publish something every single day. Then people could turn off CNBC and just tune in to what you say because, aside from it being brilliant, you have access to knowledge that isn’t in anyone else’s brain, as far as I can tell.

You’ve got this really unique combination of perspectives, and it’s incredibly valuable. Using distillation as a starting point, what founder advice would you give? We’re seeing numbers that are completely unprecedented at ages of founders that are also unprecedented.

You saw Sergey and Larry when they were very young, but now they’re even younger. We’re talking 18- and 19-year-olds now. What advice would you give them?

5. Learning Systems Create Winners

I think the most important thing to say is that the barrier to entry for starting a company is effectively zero now. Let’s think about it. What do you need to have a company? You can register it online. You need some money to get started. You have to pay yourself.

You don’t really need any programmers. You need a couple of people willing to pay Google or Claude, or what have you, to write the code for you. You’re pretty good there. You can use third-party logistics companies if you’re building a hardware device, and you can use essentially contract manufacturers to build whatever hardware device you want.

It looks to me like, for hardware and software, the barrier to entry is almost zero. That sounds great until you realize that, as a result, you’re now competing against everyone all the time.

As an example, in my career, which has spanned more than 55 years now doing this stuff, the key thing that’s true is the compression of time. The other thing I would say to founders is that it’s really important that everything you do be learned and not specified.

I’m doing a couple of startups on my own. We’ll see how well they do. With them, I say, “I know nothing; learn everything.” You can learn how to support your customers. You can learn what the customer wants. By learning, I mean it in the AI sense: learning as part of either supervised or unsupervised training.

If you take a learning approach, then you build a system that, if it works, will explode.

Mm-hmm.

Because once learning accelerates, you get into a quasi-monopoly position. The philosophy of winning goes something like this: Run as fast as you can, get there as quickly as you can, build it around learning, and if it takes off, you’ll be a hero because once it learns, it learns how to become stronger.

And eventually, in 2 or 3 years, you'll start to have various forms of reinforcement learning that are self-replicating. Then you're likely to get further accelerations. That's the most likely path for the next trillion-dollar company past the equivalents of Anthropic, OpenAI, et cetera.

When you're looking at an investment yourself, the learning loops concept that Eric just mentioned—on the Moonshots podcast that we did with Eric—he actually described in detail the 3 or 4 different types of learning loops that he looks for. It's definitely worth exploring. It's a lot more than we have time for right now, but it's really brilliant. Aside from the actual business plan and the learning loops, what do you look for in founder dynamics, the founder team, and team members?

It's always the case that the founders are really smart. They're very quick, and they're very interesting to talk to. It's also true that you need them to be able to hire a network of people like them. A simple way is that if you talk to them and they seem really interesting and really dynamic, and you find that they're in a network of such people, you're likely to have winners.

What you do is say to them, “Show me—don't tell me the product,” which, of course, is what they want to talk about. “Show me how you're going to build a system that goes from zero to infinity.” There's a lot of discussion about zero to one, and there it's complete compression. Get that thing done. But once you have one, how are you going to scale?

Our industry makes enormous wealth for the founders when you build a platform that is scaling. That's the lesson. What is the lesson to offer? Build a platform that scales. A platform is defined as something that others depend on that you provide, right? The stronger the platform, the more networked it is, the more interconnected it is, the stronger the network lock-in. That's just generally true.

It was true for Microsoft. It was true for what I did 20 years ago. It's true today. I'll give you an example. Everyone's looking at where the economics are for the large LLM companies. They don't have strong enough network lock-in yet, but you could easily imagine that they would develop it, and I'm sure they have that in the back of their mind. That's part of the reason why their valuations are so high.

We have 1 minute remaining, and I really want to use it to milk you for a quote that we can loop on our wall in the office. The perfect quote to me would be about the importance of this moment in time. So many of the founders weren't around when Microsoft could—when Bill Gates could wake up in the morning and decide, “Yeah, I'm going to destroy WordPerfect today, and I'm going to destroy Lotus 1-2-3 today,” because he just had that power at that time. Novell was certainly in those crosshairs, too.

Then there was this magical moment of the internet explosion, where companies like Google and many, many others could thrive. But after that, things got static again. Now we're in the most explosive time period for opportunity for entrepreneurs that I've ever experienced in my life. But very few people remember all the way back to the internet explosion era. I'd love to get your quote or your thoughts around the importance of this moment in time.

6. AI Defines the Next Century

I firmly believe that the arrival of nonhuman intelligence—AI intelligence—is at the level of electricity, the invention of fire, transportation, and so on in human history. We are fortunate to be living in a time of great historical consequence. The next 10 years are probably the 10 years that will have a greater determination over the next 100 years than anything before because of the inventions of these new tools, and the tools are very powerful.

Remember, they're powerful because they can equal and, in some cases, surpass human intelligence, and human intelligence is everything for a society. The countries and companies that embrace this nonhuman intelligence correctly and aggressively will be the big winners. The companies and countries that are slow or let other people do it will lose because that is the source of excellence, the source of leadership, the source of growth, the source of everything, the source of innovation—everything.

Economic growth comes from the application of intelligence to discover new things and solve new problems. We are on a huge course to accelerate that here in America, and I'm very proud to be part of it.