Dr. Eric Schmidt: The Future of Technology at 300

So first Tom, thank you for doing this. I want to thank you all for being here. This was an experiment that I came up with and I said, well, you know, it'll probably work and it's worked enormously well. And I want to thank Illy and the entire team putting this event on. We're committed to national security. We're committed to working with the government. We're committed to working with you who work with the government. We want to make America safe and powerful and keep it that way for many many decades. To answer your question, I spent lots of time in Ukraine. I have a company there. And I spent lots of time on the front lines, which is not for the faint of heart. And I think now we're seeing there. So, it's important to understand that Ukraine is the laboratory for future war. I'm not in any way endorsing this war. It's a horrible war. I've seen plenty of people die. It's just horrible, horrible, horrible, horrible. Having said that, it's important to understand what's going on. Today the front as it's known is quite wide. That's the actual line marking the two sides. And the people who are on that area are in foxholes. The minimum number of days they spend in the foxholes is 50, but in many cases it's up to three to four months. How do they survive? Drones supply them with water and food. If they stick their head out of the ground, they're killed. That's what a transparent battlefield looks like. And I don't think we Americans understand this. I think almost no country understands that that's what the future of drone war means. The result of this is that the front is largely frozen. It's very hard for either side to go against the other and people take enormous risks running along the fields, hiding under trees and so forth while the drone hunters are hunting them and it's just horrible to watch. In Ukraine, they've recently announced that they've hit their fourth month of killing their quota of Russian targets, which they say is 35,000. To give you an idea of the scale of this war, it's hard for us to understand. Think about the number of people who we lost in Vietnam in the context of what's going on in Ukraine. Ukraine numbers are not released. But it's substantially lower than 35,000. And the consequence of this is anything that moves is destroyed. So get rid of your tanks, armored personnel carriers, artillery, flamethrowers, all the fun stuff that you see in movies, right? It's sort of over. Replaced by these powerful robotic systems. Technically, the way they work is they can work without GPS. They can communicate through electronic warfare. That technology is getting stronger on both sides. So this is the future. The battle now has moved to the midstrike area which is roughly 40 to 100 kilometers and that's where the supply line on the other or the enemy side is supplying these people who are unfortunately in the foxholes and that's where the bombing is and that's where the war is largely being fought right now. And Ukraine has pretty much stopped Russia's advance, which is extraordinary, and is slowly beginning to respond successfully. You asked about AI. If you take the principle that I described, which is that basically air, land, and sea drones are going to be the way that is code for robots and automation and AI. They will all be interconnected. They will all operate in the equivalent of swarms and they'll all operate in what are called coordinated attacks. So a typical example is you'll have a target and you'll surround the target by a number of drones and you hit it simultaneously and most targets do not have the ability to simultaneously in a 360 azimuth defend themselves. So that's roughly where we are.

So it's a slightly different subject would be on AI. They're all interrelated, of course. A week ago, China released Deepseek V4. And when I was working on the National Security Commission for AI, we worked really hard with Trump one and Biden to implement hardware controls for chips. We were very successful collectively, all of us, to limit hardware access to China. And I was strongly in favor of this policy, as were many people. It looks like the policy has worked, but it's largely beginning to not work. The Chinese are clever. They're smart. And they have now built systems that are within striking range of the top models in America, but with much much less powerful hardware. The technical summary is that they use a chip called the Ascend chip. It uses a slower process and they've invented new software ways to get around the various latency and architectural problems. What I like about this is that we have real competitors. What I don't like about it is China is very focused in broad diffusion of this technology globally and it's all open source which means it's largely uncontrolled and not controlled in any way by us. So I think a fair statement is a year ago I said I thought that they were one to two years behind. It looks like they've caught up enough that the most recent analysis is China's within six months which is a nanosecond right in our world. So it gives you a sense of how committed China is to the AI leadership position and they're not going to stop. It makes you feel any better. In order to do this, you have to have a whole country of, you know, engineers, scientists, nerds, money, hardware, and so forth. There are not going to be many countries that can do this on their own. But China's clearly one. America's another one with our allies, of course. Maybe there'll be a third or fourth.

So let me say some uncomfortable facts which I think are true. And if you don't believe me, please write what I say down and then check me on a year in a year. We build an enormous number of tanks, ships, missiles, human-driven jets, etc. They're all now vulnerable. Anything that's exquisite, big, slow, and emits heat, it's a problem. So the conclusion you come to if you sit in my position is we should be focusing our national security on automation of land, sea and air and so forth. We have fantastic military, very very smart people. We should be able to make this transition. It takes a while. And by the way it doesn't produce less jobs. It just produces different ones. And by the way, this doesn't mean the budgets go down. They probably go up, but they go up with different things. So, for example, why would you build artillery for a 155 piece? These big 155 things which are completely untargeted and only go about 15 km. Makes no sense. You should be building the equivalent thing for a drone, which is far more, doesn't cost any more. In fact, costs probably less. You can get a lot more for the same amount of budget. Plus, you can swarm them and so forth and so on. So as a matter of military doctrine, I think all that stuff that we learned watching the courage of the Marines in Iwo Jima, all of that stuff, it's all true. And it's also over, right? That the future combat will be largely robotic. It will be automated. It will be controlled by the laws of war. America has a very clear model of human control, which I strongly support. I helped write it and I feel very very strongly that this is the direction that the nation should take. With respect to AI in general, I'll just confess what's really going on in my mind. I'm in mourning and I'm in mourning because I started as a programmer when I was basically 13 or 14 and that is essentially over, right? How can your entire identity and life's career as a programmer computer scientist be over in one lifetime? It's supposed to be like your children or your grandchildren. And what's happened, I'll give you an example. So the way programming is done at the state-of-the-art now is a programmer wakes up, goes to the office because they're social and they sit there but they don't have that many friends. So they sit in their office and they collect 10 Claude friends or 10 Gemini friends. And they set them up with objective functions and they watch what they wrote, the code that they write. Then they go out to lunch and they make sure that they have long enough tasks that they continue working while they're having lunch. So then they go back to the office and they do this and then it's time to go home to see the family and so forth and they set up an objective function of this is what I want you to do and here's the verification and it's a long enough project that it takes them all night. So, they've gone from writing code, which is what I did, and I thought I was the oddest thing ever when I was like 20, and I'm sure that wasn't true, but welcome to being 20, to now having this kind of power. And I've watched what these systems can do, and it blows me away. Right? So, we have a major major change happening. This started about October of last year. It's new for this conference. And if you're writing code in any traditional way, stop, right? It's over. And go. And if you manage a company and you have software programmers say, why are you still writing code the way you did it six months ago? This has a lot of implications globally because it means that the productivity of software which has always been an issue is going to become far far greater. It also means that individuals will be able to build incredibly powerful applications and so forth. You saw for example this huge selloff in software as a service stocks essentially traditional software stocks. That's because the market decided that those products would literally go away which of course is not true. They're probably oversold but they're directionally in trouble and they themselves will have to adapt.

So what it says is that the traditional writing code, you know, like if then else, the stuff that I used to do, you're not going to do that anymore. What you're going to have to learn is to write down what you want in a way that the computer can understand it and then monitor it producing it. Do you remember two or three years ago there was this whole thing about hallucinations and we couldn't trust these systems and they were off, you know, falling in love with the therapist and things like that. That's largely under control. I can sit in a meeting and somebody will say well we just don't have the right user interface and I can basically say while the meeting is going on I can specify within these tools build me a UI with these features and it produces that. Doesn't work yet but that's the change. So you learn the new programming if you will is learning how to manage the programming tools. You've gone from building the house to becoming the architect and the builder is under your control.

And I should say, by the way, that the people who were, it's always been true, sorry to go on about this. It's always been true in software that the very tippy top programmer was worth 10 times what the next level was. It's similar in if you think about football salaries or NBA salaries, their productivity was so great. So my guess is that there will be a smaller number of software people in aggregate but the very very top ones, the top ones that can manage this productivity will have enormous economic return to themselves, their corporations, things like that.

So, again, let me give a bit of the background. AI went from language to language to language to action. So when you go and look at the robots that we have here in the demo that you can communicate with them, they look at something you can communicate with them and they take an action. Those are called VLA models. That technology is emerging very very fast. Those require an enormous amount of training data. The reason language to language works so well is because there was so much training data and we sucked in all the world's language and built all these incredibly powerful models. And one of the emergent behaviors of these things is they could begin to do planning. So somehow they learned how to do planning. So you can say, again it's called chain of thought reasoning. You can say show me the steps to do something. So all of that is pretty well assembled. The next thing is prediction. And these models are very good at predicting things. So in war for example, you do predictive battlefield analysis. I want to know what happens if I do this. What happens? Very reasonable. The same is true in a company or government or whatever. It's not unique to the military. The next one is called reinforcement learning and that's a simulation task and that's how we for example won the game of Go at Google which was thought to be incomputable. That's how something called AlphaFold was done at Google which everyone won Nobel prizes for that. The ability to understand how proteins fold is a very big deal. This technology is generally available now if you have the data. So if you want to do the kind of stuff you're talking about for national security reasons you have to have a great deal of training data. The only place I have found that has enough training data is inside of Ukraine. And shockingly, the Ukrainians will not release it for obvious reasons. I'm joking. It's highly classified. You can't get to it for all sorts of very good national security reasons. So, for America to win in this particular area, we're going to have to have lots of training data. One way to get that is by doing things like we have in the cage outside where you basically have drone competitions. And I think my guess is that our military ranges will ultimately have drone on drone competitions with live fire and that's how the training will occur.

So, just one more thing on energy for the national security. So, our military has now, which is great, has now figured out they need real classified data centers and so they've started looking at how to partner with the commercial sector and so forth in this way. And I think this is wonderful. We're not going to be safe without our government having its own very very powerful data center, something which they've never really had. And when I was working for the Secretary of Defense, and we're on the Ash Carter stage, so I'll always owe Ash a lot for getting me involved with this. There was essentially none. So, the military now has a plan to do this. This is very welcome. In general, you're going to read, let's go through the sort of what you'll read in the press. The AI boom has contributed between one and 2% of the GDP of America. That's a very big number. For most of you, you're going to say, 'My god, these people are idiots.' I mean, how could you possibly waste all that money? That's not my view. I think that the AI revolution is underhyped, not overhyped. This is my opinion. And I started one of a number of data center businesses to study this and to build data centers. And I can tell you roughly what you conclude. First, the interconnection grid is not going to survive all of this new power. There's not enough switches. There's not enough wires. There's not enough management. The interconnection queues as they're called are very long. So, what people are now doing is they're building data centers in remote places with their own power plants. The technical term for that is behind the meter. What they do is they build essentially air cooled data centers of enormous scale. You've seen the pictures. They're essentially air cooled systems. They don't use as much water as you think. The only water they use is internal for internal transfer. Very little water escapes, but they use a lot of electricity and they're measured in megawatts. The largest data centers are in the one and two gigawatt stage and there are people talking about a gigawatt data center costs something on the order of 50 billion with a B number. How could these people in my industry be so crazy to raise all that money and build all the data centers? And the answer is they see the demand. And my industry has changed so that if you want to make more money you need more servers because the servers are how you make the money because that's the reasoning. So, absent an AI breakthrough algorithmically, which is a change to what is called the transformer architecture, which was invented at Google, thank you, in 2017, we're sort of stuck with this cost structure. So, my own opinion is you're going to see a lot of remote America build these behind the meter power plants, it takes forever to get the turbines and all of that. And I think the real limit to AI is not energy. It's actually cash. Because when you add up the cost of these, if you take 50, I'm just round numbers, 50 billion a gigawatt, so 10 gigawatts is a half a trillion dollars. How many companies, countries, and so forth can hand an industry a trillion dollars of capital? Very, very few. The Chinese could certainly do it. I don't know if they're doing it. I'm going to try to find out. In America, there are people who hope that's going to happen. It's interesting that you can finance these things because the brilliance of the American capital market allows us to borrow that kind of money. So, for example, the Europeans can't do this, right, which they're sort of sore about, but this is good for America.

Well, you know, Henry was my closest friend and so I miss him, you know, every day and I don't know how to bring him back. The empty chair is for his. Yes, he would be very helpful now in the Iran crisis, for example. He understood all that very well. And Henry was very concerned about competition between America and China in AI that it would somehow unleash an unstable situation. And he felt very strongly and started a so-called track two dialogue which continues with the Chinese. And the concerns have, I'll give you an example if you look at Mythos which is the Anthropic model that's so powerful, we collectively believe that Mythos is just the first of a set of very powerful models that have the ability to for example do cyber attacks and this is not a desirable thing. It's desirable if you're in offensive cyber attack mode, but it's certainly not desirable if you're a company or in the military and so forth. And so you have to think of it in the following way. You have these enormously powerful models coming out. How many will there be? Well, there's four or five in the US. There's at least three powerful ones in China. There'll be a few others. Let's say there's 10. How do you want to organize the world? Can you have meetings where all the players agree to limitations? And the concern goes something like this. You and I are on opposite sides, which we would never be, and we agree. But then I build something which you know I'm doing, and it's really good, and I'm going to get a significant advantage by undercutting you, by running ahead of you, by violating our agreement. It sounds a lot like OPEC, right? So we don't have language among the countries yet on how to handle this. The Trump administration has figured this out and wants to start working on this which I think is very welcome and we'll see. The general issue of global instability is actually even more serious than that. Mythos is just a good example that everybody's familiar with. But the Chinese models are open-source, open weights. So for example, this means that if you want to take DeepSeek V4 and you want to add powerful capabilities and you're evil, which none of you are, and I'm not suggesting the Chinese would do this, but somebody could do this. Some evil person exists in the world. How would we know that? How would we police it? How would we see it? At least with the American models, the big ones are under corporate control, proprietary. We know who to call, the military can call, the police can call, and so forth. But if you release this stuff to open source, it's dangerous enough that a lone actor, a new kind of terrorist could emerge. Obviously, we want to avoid that and we want to acknowledge it's possible and we want to get ahead of it and it requires multistakeholder agreement.