Dr. Eric Schmidt: The Future of Technology and AI

First of all, Tom, thank you so much. I'm grateful that you came to participate, and I want to thank everyone who came here today. This was an experiment Uri conceived, and I thought it would work out well. It really turned out wonderfully. I'm truly grateful to Uri and the entire team that supported this event. We are seriously committed to national security and intend to cooperate with the government, and I want to continue working with everyone already doing that. We want to keep America safe and strong for decades. To answer your question, I spend a lot of time in Ukraine—I have a company there and visit the front lines often. Honestly, it takes a strong heart to be there. We are seeing firsthand what's happening. Ukraine has become the testing ground for future wars. I absolutely do not endorse this war—it's truly terrible. I've seen so many people die. It's truly horrible. But it's important to understand what's happening. The so-called front line spans a very wide area—it's the actual line separating the two armies, and people there are holed up in dugouts. You spend at least 55 days in a dugout, often three to four months. How do you survive? Drones deliver water and food. If you stick your head above ground even slightly, you get shot immediately. This is a transparent battlefield. I don't think Americans understand this reality, and most other countries don't either. This is the future of drone warfare. As a result, the front line has barely moved. Neither side can easily advance. People take enormous risks running across fields or hiding under trees. It's heartbreaking to watch while drone hunters shoot down drones. Ukraine recently announced they've been killing Russian targets at a pace now entering the fourth month, totaling 35,000. It's hard to imagine the scale. Compare the number we lost in Vietnam to what's happening in Ukraine. Ukraine's numbers aren't public but are said to be considerably less than 35,000. And everything that moves in this war gets destroyed. Tanks, infantry fighting vehicles, artillery—the exciting weapons from movies are no longer usable. They're being replaced by powerful robotic systems. These drones can operate without GPS and communicate via electronic signals. Both sides are rapidly advancing this technology. This is the future of warfare. The main battlefield has shifted to mid-range attack zones, roughly 40 to 100 kilometers—where enemy supply lines deliver materials to dugout soldiers and where bombing occurs. Most of today's war is fought there. Ukraine has nearly stopped Russia's new forces, which is truly remarkable, and they're beginning to counterattack. There was a question about AI. The principles I described—drones in air, land, and sea becoming the main force—mean robotics, automation, and AI become central. These will all be interconnected, coordinating like swarms for synchronized attacks. If there's a target, multiple drones surround it and attack simultaneously. Most targets cannot defend against simultaneous 360-degree attacks. That's largely the situation today.

Of course it's all connected, but last week China released DeepSeek V4. When I worked on the National Security Commission on AI, I worked hard with both the Trump and Biden administrations to introduce chip hardware restrictions. Working together, I think we were quite successful in limiting China's hardware access. I personally strongly supported this policy, and many others did too. The policy seemed to be working, but recently it's become less effective. The Chinese are incredibly clever—they're now building systems close to America's top models but on much lower-performance hardware. They're using something called Ascend chips, a slower process, but overcoming delays and architectural issues through clever new software techniques. What I like is that a real competitor has emerged. What I don't like is that China is investing heavily in global proliferation of this technology, and it's all open source—meaning it's mostly uncontrolled. To be honest, a year ago I thought China was one to two years behind. But current analysis suggests they're catching up within six months. In our industry, six months is an instant. This shows how serious China is about AI leadership, and they definitely won't stop. If there's any comfort, achieving this requires an entire nation's engineers, scientists, geeks, funding, hardware—everything. Not many countries can do this independently. China certainly can, the US certainly can, and including our allies, maybe a third or fourth country could emerge.

Let me share something that might be hard to hear but I believe is true. If you can't believe it now, write it down and check back in a year. We've built enormous numbers of tanks, ships, missiles, and manned jets, but now all of those have become vulnerable. Anything large, slow, and heat-emitting becomes a problem. So national security should invest much more in automation across land, sea, and air. Our military is outstanding with truly excellent people—they'll overcome this change, though it will take some time. This doesn't mean jobs decrease—it means the nature of work changes. It doesn't mean budgets shrink either; budgets may increase in different areas. Why continue building large 155mm artillery shells? Those can't aim properly and have a range of only about 15 kilometers. Honestly, they're pointless. It would be far more efficient to build drone-based equivalents at lower cost. You get more for the same budget, and drones can operate in swarms. The military wisdom we learned from the courage of the Marines at places like Iwo Jima—that's all real. But the era has changed. Future combat will be mostly robotic, automated, and controlled according to the laws of war. America has a clear model of human oversight, and I strongly support it. I helped write those rules, and I firmly believe this is the direction AI should take nationally. To be honest, let me share what's in my heart. I'm in a state of grief. I started my career as a programmer at 13 or 14, and it was my life's identity. But that's coming to an end. Having your life's work end in just one generation—normally you'd expect it to continue through your children and grandchildren. The cutting-edge way of programming today: a programmer wakes up, goes to the office, gathers about ten AI friends like Claude or Gemini, sets objectives for each, and watches what they write. Then they go to lunch, leaving the AIs with longer tasks. After lunch, same thing. When it's time to go home, they tell the AIs what to do overnight and check results in the morning. So the work I used to do—writing code myself—has shifted to delegating to AI. When I was 20, I also thought I was the most different person, but probably not. But now these tools have this power. This is what it's like in your twenties now. Seeing what these systems can actually do is truly amazing. A huge change started around last October—it's a new topic even at this conference. If you're still coding the old way, stop. The era has changed. If you're running a company and your programmers are still coding the same way as six months ago, you should ask why. This has major global implications—software productivity, always a challenge, is about to increase dramatically. Individuals will be able to build incredibly powerful applications. You've probably seen the big sell-off in software-as-a-service stocks and traditional software stocks—the market judged these products would disappear. Of course that's not really the case; they were probably oversold. But the direction is tough and they'll need to change.

The point is, the old way of writing conventional if-then-else code is no longer needed. From now on, you write out what you want in a form the computer can understand, then watch over it and verify the results. Remember two years ago when hallucinations were a big topic and people said these systems couldn't be trusted? That's now largely under control. For example, during a meeting someone might say 'the user interface isn't quite right.' If I specify what kind of UI I want using a tool, it gets built automatically. It's not perfect yet, but that change is happening. The new programming is about mastering programming tools. Until now you were building the house yourself—now you become the architect, and the builder follows your instructions.

And I'd like to add—it's always been like that. Sorry, this is getting long, but in the software world, it's always been said that the very top programmers are worth about ten times the next level. Think about soccer players or NBA salaries—it's a similar dynamic. Their productivity is incredibly high. So my prediction: while the total number of people in software may decrease, the truly top people—those who can control this productivity—will generate enormous economic returns for themselves and their companies.

Let me provide some background. AI has evolved from language to language, and then from language to action. The robots you saw in demonstrations—we can communicate with them, they observe something and take action. These are called VLA models. The pace of evolution is truly remarkable, but these models require enormous training data. Language-to-language AI succeeded because there was vast language data worldwide to create incredibly powerful models. One new characteristic is planning—the ability to make plans. AI has learned chain-of-thought reasoning—ask it to show the steps for doing something, and it explains step by step. These parts are well developed. The next stage is prediction—these models excel at it. On the battlefield, you'd want predictive analysis of what happens if you do something. This is logical and applies beyond the military to business and government. Beyond that is reinforcement learning, which involves simulation. Google's acquisition of DeepMind was driven by this—Go was considered computationally impossible but they mastered it. Then there's AlphaFold, which made incredible progress understanding protein folding. People won Nobel Prizes. These technologies are now available to anyone with enough data. For national security, you need enormous training data. The only place I've found with sufficient data is inside Ukraine, but Ukrainians won't share it—it's classified. So for America to win, we need to gather lots of training data. One way is holding drone competitions in the field. My prediction is that eventually we'll have drone-versus-drone competitions with real targets on military training grounds. That's how we'll collect training data.

One more thing about energy for national security. Our military has begun to realize—they need proper classified data centers. They're starting to think about collaborating with the private sector, and I think this is truly positive. Until the government has its own incredibly powerful data centers, we won't be safe. We've never really had such things. When I worked under Defense Secretary Ash Carter—who gave me my start in this field, and I'm truly grateful—at that time there was almost nothing. Now the military has plans to realize this. This is very welcome.

Generally, as you may have seen in the news, there are claims that the AI boom is contributing 1 to 2 percent of US GDP. That's a huge number for many people. They might think, 'Are these people crazy? How can they waste so much money?' But I disagree. I believe the AI revolution is not just appropriately assessed but actually underestimated. This is just my opinion, but I've researched this field and built several data center businesses, so I can speak from that experience.

First, the current grid cannot handle the new power demand that will be needed. There aren't enough switches, cables, or management capacity. Interconnection queues are extremely long. So what everyone is doing now is building data centers in remote locations with their own dedicated power plants. In technical terms, this is called 'behind the meter.' Basically, you're building massive air-cooled data centers. Some of you may have seen photos, but they're fundamentally air-cooled systems. They don't use as much water as you'd think — only a tiny amount internally. But they use an enormous amount of electricity, measured in megawatts. The largest data centers are 1GW or even 2GW scale.

A 1GW data center costs roughly five trillion yen to build. So why does my industry raise that kind of money to build them? The answer is: because there's demand. The industry model is also changing — if you want to make more money, you just need more servers. The more servers, the more profit. Unless there's a major algorithmic breakthrough — for example, the Transformer architecture, which Google invented in 2017 — we can't escape this cost structure. So in my personal opinion, we'll see more and more behind-the-meter power plants built across rural America. Arranging turbines takes time, of course.

And I believe the real constraint on AI isn't energy — it's money. As I mentioned, 1GW costs five trillion yen, 10GW costs fifty trillion. How many companies or countries can provide that kind of capital to the industry? Very few. China probably could, and I'd like to look into whether they actually are. There are also people in the US hoping for this. What's interesting is that the ability to fund these projects comes from the strength of American capital markets — being able to borrow that scale of money. For example, in Europe, this isn't possible. Europeans seem a bit frustrated about it, but it's a good thing for America.

Yes. Henry was my closest friend. I really miss him every day. I don't know how to get him back. That empty chair over there is for him. He was truly someone you could rely on. For example, during the Iran crisis, he understood the situation very well. He was also very concerned that AI competition between the US and China could create instability. So he strongly pushed for what's called the Track Two dialogue, which continues on the Chinese side. And that concern persists.

For example, if you look at Anthropic's Claude model, it's extremely powerful. We all think Claude is the first in a series of very powerful models to come. These models have capabilities like carrying out cyberattacks, which isn't desirable. It might benefit the attacker, but for those on the defense side — corporations, militaries — it's completely undesirable. So we need to think about this: very powerful models are emerging one after another. How many will there be? The US has several, China has at least two strong models, and others will emerge. Let's say there are several.

So how should we organize the world? Can we hold a meeting where all stakeholders agree on restrictions? The concern is like this: imagine we're on opposing sides — though in reality we're not — and we reach an agreement. But then I'm building something new, and you know about it, and it's truly superior, so I break the agreement and advance ahead of you to gain a big advantage. It's like OPEC. Right now, countries don't have a common language or rules for dealing with these issues.

The Trump administration is aware of this problem and wants to start working on it. I think that's very welcome. Let's see what happens. Actually, the problem of global instability is even more serious than this. Claude is just one well-known example. But China's model is open source with weights published. So for example, someone could take DeepSeek and add even more powerful capabilities. If someone with malicious intent — of course not you, and I don't think China would do this, but the possibility isn't zero. There are bad actors in the world. How do you find them, monitor them, prevent them? It's difficult.

At least with major American models, corporations manage them, you know who to contact, and the military or police can respond quickly. But when these things are published as open source, a single malicious person — a new type of terrorist — could emerge. We absolutely want to avoid that, and we want to acknowledge the reality that this could happen and take preemptive action. For that, many stakeholders need to reach agreement.

There are a lot of people in the tech industry who talk grandly about this, but many haven't actually studied economics or social sciences. They were too busy programming to take those courses. So first, I think you should ask people who have actually studied these fields. When I talk to labor economists, their argument goes like this: AI's efficiency gains are enormous, so companies will definitely adopt AI. That makes companies more profitable and bigger. Of course, there will be disruption and pain along the way, but ultimately companies grow and the entire economy expands. This has been the case for the past 200 years and hasn't changed.

So the first question is whether this pattern will continue, and I believe it will. And I think we'll actually face labor shortages, because productivity will rise and the economy will grow at incredible speed. That's the optimistic view. The accelerationist view goes even further — that things will get so cheap everyone can relax on the beach. But my own thinking, and this is speculation, is that probably won't happen. The reason is that at least in the American system, people love competition — lawyers love fighting each other, people love cheering for sports teams. So it won't be everyone sipping margaritas on the beach.

Instead, people will engage in higher-level competition, and that will generate sufficient employment. Actually, many people blame AI for job losses, but if you look at the data, only two categories are actually shrinking. One is young software engineers in my field — for the reasons I mentioned, it's harder for them to find jobs. The other is low-wage customer service work. And while these may increase in ten years, it's not progressing as rapidly as people think. I haven't spent much time worrying about this issue.

Thank you. You're exactly right, and thank you for raising this point. In our book, Dr. Kissinger and I wrote that AI is too important to leave to technical people like me alone. We really need more philosophers, humanists, and what you might call humanities scholars. Recently, those people are starting to get involved. The reason is that humanities researchers are training graduate students, and currently the most interesting topic for humanities graduate students is how powerful tools like AI will affect their fields — psychology, political science, governance, and so on. So the situation is gradually improving.

Companies use the term 'alignment' or 'harmony with humans' to refer to both values and safety. First, there's the safety mechanism — for example, which AI model, if you tell it 'I want to kill myself,' responds 'Please don't. Contact the suicide prevention hotline.' That was added later. AI had learned methods of suicide, but those answers are now suppressed. That's a good thing. The more subtle issue is the style, communication, and values of the AI you're conversing with, which are shaped by training. The question is: do we really agree on human values?

Probably everyone here supports Western liberal values — women's rights, minority rights, freedom of expression. But what if a Chinese AI model is trained with different values and that spreads? With open source, it's entirely possible that most AI models worldwide become Chinese rather than American, since open source is cheap. Of course, AI with American values is also being built — Google's Gemini clearly has American values, and NVIDIA's Nemotron is similar. There are also startups building better things, and I'm very hopeful about that. But when it comes to human values and alignment, we have to choose which values. I like AI with American values.

Francis Lee wrote a book about the aftermath of COVID. She says people were gaslit during COVID, which is why they lost trust in science and news. I understand that feeling well. Let's start with the basic facts. This is Washington, so we have to start with facts.

America has remained a world leader for hundreds of years because of scientific and technological innovation. The innovation America has led has made us exceptional. That's why it's so important that science remains neutral, fair, well-regulated, and adequately funded. Personally, I've been fortunate to receive a lot of money from Google, and I intend to return that to science in some form. Our national security depends on science, and the future that AI will open up depends on science. So first, I want to emphasize how important science is.

Now, regarding the specific question about misinformation — this isn't solved yet. Think about it carefully: people tend to engage more with frightening information than with long, rational information. This throws off the economic balance. In other words, the CEO of the service provider doesn't necessarily have an incentive to tell only the truth. This dynamic exists in every democracy. But the most important thing I want to convey to everyone here, regardless of political affiliation, is that if we want to make this country better and grow it, we must invest in the fields we've been discussing.

For example, over the next five to ten years, applying AI to science and medicine will produce truly remarkable results for America. These studies are mainly conducted by graduate students and postdocs under faculty guidance. The ones actually producing new discoveries are these young generations. There will be amazing breakthroughs in energy, materials science, and molecular bonding. If I said everything you buy will soon have twice the performance at half the price, you'd be happy, right? That's within our reach. And I haven't even mentioned medical advances and new drugs yet.

I was talking to Demis Hassabis, and he said his company aims to solve all diseases within twenty to thirty years. That's an incredible goal. I want to be part of it too. And what do we need to do? Invest in science. Thank you very much. I think there's time for maybe one more short question.

What a perfect closing question. Let's think about seventeen or eighteen year olds. Many of you here have children or perhaps grandchildren. Their lives will be shaped enormously by this new form of artificial intelligence. AI will be a friend, a counselor, an advisor, and a programmer. As I mentioned, my career is essentially over. So the most important thing for them is teaching how to use these tools.

My specific proposal is that every university nationwide should make learning how to use these tools a required first-year course. And I will design that course. It starts with learning how to harness AI's power and ends with students building something meaningful — like a system to evaluate their friends and make themselves look good. That's a meaningful project for an eighteen year old. Some might say this should be in high school too. Sure, but let's start at universities and then expand to high schools for those who missed it.

Let me give one example. In my work with the military, when I hired someone, my only rule was that they send me a photo of themselves at graduation with their parents. That was the promise. He was extremely talented — he skipped every normal step and is now working on metacognition and swarming research. This is the future. I want to find people like this and help them succeed early. The power of these tools is truly remarkable, and it should significantly address the dissatisfaction young people feel about life and economic opportunity. And it's fun. That's actually why I got interested in this field when I was young.

By the way, I grew up in Virginia. My father bought me a teletype, and I became obsessed. When you're young, building things is fun. Today's AI is an incredibly powerful tool for creation — you can design almost anything. And yes, that includes bad things too, so we also need to teach people not to do that. But universities are still stuck in old teaching methods and reluctant to change. However, that approach will soon be over.

The era is coming when teachers will teach students how to ask creative questions — not just memorize. By the way, when I was in Virginia in seventh grade, I had to memorize all the counties in Virginia — about fifty of them. I managed to learn them all. But why does that kind of thinking still persist today? These systems are available anytime, incredibly powerful, with amazing analytical capability. What's needed is imagination and careful guidance. That's exactly what we should be teaching. Thank you very much for everything.

Of course. Thank you. That's very kind of you. Thank you.