Transcription
We are honored to welcome Dr. Eric Schmidt, chair of the Special Competitive Studies Project, for a fireside chat with Jean Mizer, host of the Special Competitive Studies Projects Natk Tech podcast.
Dr. Schmidt: Yes, great to have you with us.
Thank you. It's great to be back here. I have a little—I was going to praise Illy, but then I realized he only gave you all 15 minutes of food. That seems a bit—here and eat; they can go back for seconds, right? Except your words are going to be so enthralling, the ones who stay in—I think food is better. I do want to thank, um, Illy and the entire team for putting this on. Um, you, Illy, by virtue of your hard work and studying what Kissinger did with SCSP, you really established the structure, and it's—what I like about it is it's repeatable and scalable structure, which is how one should think about AI and energy and everything else. So well done. Thank you, Illy, and entire team.
So the MIT Technology Review had a very catchy title on an article a couple of months ago called "AI is an energy hog." Um, we're going to get to that conundrum in a moment, but first, I want to talk to you about AI and its benefits. We'll get to the costs and challenges. So what is AI going to look like, let's say in two years, as compared to what we have on our hands right now? So, um, first, there's a lot of conventional wisdom about AI, which I'll just repeat by reference: huge improvements in health care, huge improvements in education, huge improvements in science. If I look at the biggest impact in AI aside from screwing up democracy because of misinformation, which is a separate discussion, um, it looks like the fundamental acceleration in science, energy systems, material systems, and so forth is amazing, and it will continue.
Um, one way to think about the AI that you all know is think of it as language to language. You ask a question; the answer comes back. You ask a question; you can even write code nowadays. The models are multimodal, so, for example, you can take a picture and say, "Tell me what's in this picture." Uh, technically, there are APIs which allow uh one firm to call an OpenAI or Gemini API or Anthropic, for etc., and do the classification of the picture and so forth. These are all tactics that increase the intelligence of the underlying system. There are three things going on right now, this year—so less than the time frame you gave—are really interesting. One is called infinite context windows. Infinite context windows means that you can keep feeding the answer back in as the question, so it allows you to do step-by-step planning. You know, "How do I uh how do I build a house?" Well, the first is I have to find a contractor. I found a contractor; what do I have to talk to them? Then I have to have an architect. How do I find an architect? Then I have to tell the architect what to do and then design me the house. I'll give it to the architect; he can redesign it. You know, there's a series of steps.
Uh, the next one are called agents, and agents is a generally overused term, and most people think that agents will essentially act as memory sources. So an agent can be understood as it's watching something, and when it sees it, it takes an action. If it does that by knowing what to do based on what it's seen, the specs for how agents work are completely undefined in the industry. The dominant companies want to have their own agents, and they don't want the agents to interact because they want control, for obvious reasons. Many people think that there will be an agent store that you will download, might like we see with apps, but not this year. And the third one is text-to-text-to-code. Now, I don't know about you all, but I've programmed and managed programmers for more than 40 years, and they never do what I want. So can you imagine if the computer—you said, "Write me a program to do this," and it actually writes the code? In our case, uh, the program would be: search through all the literature, find out who is working on energy policy, who has a technological background or a role in which they have to be technologically literate, identify them, rank them, score them based on whatever our goal is, um, and and then automatically invite—send them an invitation. If they say yes, say "Congratulations." If they say no, "Why not?" and call them and with a synthetic voice tell them that they're idiots for not coming. That's the kind of program I would write. Thank God I'm not doing that, but but you see how easy it would be to automate tasks. So that's I think the first step. The next step is not as clear. There are uh there's sort of huge contests uh there's a huge set of contests going right now which are at a scale that's unimaginable. You have the big three in the US: Anthropic, which is allied with Amazon; Gemini, obviously from Google; OpenAI and Microsoft. And let's assume they all do really well. It looks like they're doing really well. I can talk about what their problems are, but fundamentally they're they're doing well. You have Facebook, which has chosen an open-source path for the 400 billion model, that has a lot of implications strategically, right, which we can discuss. All of these are vying for the best reasoning, the best answers, and then the best predictive analytics, the best image classifiers, the best multimodal—that technology then diffuses, or the technical term is distilled, into more specialized models, and I think that's the action you'll see in the next one to two years.
You did not mention artificial general intelligence. First, for those of us who aren't necessarily um totally up to speed on AI, what is it, and where are we? There are multiple definitions of AGI, but the—it's the term has been around for 15 years. The basic idea is: what is the point where you have the flexibility of a human in your intelligent system? So one way to understand it today is that we—these are called narrow AI approaches, although they're not certainly not narrow—you basically—they're they're initiated by a human. At what point is the question: can the computer generate its own objective function, its own goal, and how will that emerge? Uh, the—there's what I call the San Francisco school, because they're all in San Francisco, uh, which is a separate set of issues, and they all talk to each other, and they've all convinced themselves that within two to three cranks of the system—crank is about 18 months—you get to AGI, and they define AGI as intelligence greater than the sum of human intelligence. I personally think that that's likely, but not in 3 years. Not in—what is the time frame? Do you—we don't know. And and the work I've been doing recently looks like these things occur in waves, and the waves eventually concatenate, so it's combinatorial innovation. So an example to me, the most interesting one is math, and I'll explain in a sec. Everyone's focused on everything else, but here's why math is important. There's a standard called Lean, which I have been one of the funders of. It's essentially a language in which case you can generate conjectures and then prove them. So you generate into Lean, and you prove from Lean, and in the situation where the computer can start generating hypotheses and generating new math, and other computers can then prove them, that's a very big breakthrough, because that's the first really flexible part of intelligence, and then you move up to other things and so forth. Uh, I don't think—to give you a a shorter answer than a long one—AGI does not arrive on a Tuesday; it sneaks up on you, and all of a sudden the systems get much more automated, much more flexible. So, for example, my my favorite example is in TikTok, of which I'm not a fan, given my age. Um, all of these people—perhaps their nationality—but that's just a gift—that's my age; it's just—you know, it's important to recognize when you're too old for something. But the important thing about about TikTok is humans spend an enormous amount of time making the videos, trying to figure out how to make them viral. Well, surely the computer understands its own virality. Why can't you just automate that, right? So I'm I'm now on a kick where I sit to myself, for everything I do: why does the computer not automatically do this for me? And I think that's the notion of an assistance; it's much more powerful than, "Oh, oh, you know, help book my flights." This is what everybody uses because that's a common one. It's really: how do I approach my day? How do my systems work? How do I get from here to there? Arrange it all for me, and I don't have to think about it.
Several of the speakers here today have talked about the impact of AI on the US economy, uh, also on US security. Could you give us your thoughts on both of those? Um, so AI in in the in the American system—AI will make our businesses more efficient, and people don't don't understand economics very well, but if you make businesses more efficient, you usually get more higher-paying jobs, and the jobs you lose are the ones which are more easily replaced. So although there's all these articles about software people going away, the fact of the matter is it looks like there's a lot of demand for more software, and even with these tools, there'll be more software jobs. But if I worked in a call center, especially in a a distant country of low low wages and so forth, I would really worry about my job, because these systems are so good now that not only can you talk to them and they talk back—whatever whatever they are—but they're they provide a better experience than someone who's in a call center, well, because you get the voice of Judi Dench, well, or or yours, which is even more pleasing.
Oh, well, thank you. But but but my my point here is is that the job dislocation that everybody's worried about is just fundamentally false; it's just not correct. And I'm tired—I've been hearing it for 10 years, and it's not—it's not proven; it's not historically true. The evidence is that automation in in general forms gets rid of the most dangerous and the most sort of unpleasant jobs and the ones which require the least amount of higher-order work. So, for example, security guards should be replaced—they're very nice, by the way—but they should lose their jobs to more automated security. But those people don't have the training to be the software engineer. Welcome to America, right? Uh, this is not a new thing; there are they're enormous amounts of job changes and every everywhere. Don't you remember that 5 years ago we were going to have all those truck drivers going to lose their jobs? The number one job that's the hardest to fill right now is truck drivers, because truck drivers—the truck driving job—is very hard; it's very complicated; it's very hard to replace truck drivers by robots. Shocking. Think about it for a while, right? When we're—how many of you arrived here in a self-driving car? My guess is zero, or maybe one—one person who is having fun. Um, you you see, and these—the technology has been around for 30 years anyway.
Going back to National Security, the US is in a very strong position, and I I think this is a credit to all the people who built the country, all of the bankers who are willing to give billions of dollars to products that have no revenue, no plan, with crazy people who look sort of like me but younger, um, and that system is the envy of everybody. It's the universities, it's the faculty, it's the at-risk capital, it's the venture—we all work together on this. It's interesting that DC is obsessed with China, and I am not. I was—now, when I look at China, I look at a country that is not doing very well on some of the core things that China needs to do as a country, and the foreign—the investment in—the valuations into in AI have dropped by a factor of five. They can't get foreign investment now into China from other countries; everybody's nervous. I'm part of a Track 2 dialogue where we talk about it. They're not—they're trailing us by a significant degree, even though she has said we are going to lead in this particular area. I say that in my company, too. However, maybe you should actually have a plan to do it. I'm not trying to give President Xi some advice, um, but the—I'm sure he'd welcome it. How—oh, she—I think that's a bad idea on my part, um, but if you look at the decisions that the Trump Administration and the Biden Administration made about restricting chips has had a deleterious effect on their progress. There are many many workarounds; there's a certain number of A100s that are still in use in China, um, there's various ways of getting around the restrictions, and it's turbocharged their research to develop their own, of course. But one of the things that happened five years ago was everyone was convinced that AI was about data; it's not right. Everybody said China will have this huge advantage because they have so much data; they have any privacy rules. That's not correct. What is correct is AI is about the rate at which you can put incredibly sophisticated STEM people to work on these problems: physicists, biologists, math people—stuff is hard. I mean, I speak as a computer scientist; some of these algorithms are too hard for computer scientists to understand; it takes mathematicians. That's how advanced this is; that's where America is strongest. The worry is that China is producing a large number of these people, too, and we have the usual debate in DC, which is never resolved: why don't we have all those people in the US? Uh, I don't need to tell this audience why—would we take the top smartest people in the world and kick out—kick them out of our country because we won't give them a visa? It's the stupidest policy on a long list out of DC, right? And it's it just never gets resolved. Our industry needs immigrants—are very very high-skilled immigrants—as well as a huge focus on US STEM training, um, which people work to.
So I'm curious, you mentioned that your view on China has changed. Was there some specific moment or some particular development that tipped you? So so Dr. Kissinger and I went there last year on his last trip before his death, and we talked to all the senior people in the—for obvious reasons involving him—and you could feel—you could feel it coming off; it was fundamentally a post-co situation, and the country's got—I mean, these are very very—I happen to like the Chinese—the citizens; uh, these are very clever people; they work really hard; they've got a lot of good things going for them, but these macro things are really a problem. You see them now dumping across the world because the manufacturing markets—they're actually—they're actually selling products that lower—lower their cost, and their cost—the best in the world—all of these kind of things. Um, so at the moment—I think this is the important part about AI and AI on—is this is ours to win now. China, by the way—there must be people in the room here understand this—have a huge program around uh fission uh fusion; they have research in fusion and so forth and so on. There's every reason to think that China understands this. One of the things that's interesting when you look at the data centers—the data centers need constant base load power. In the US, that's roughly 3% of the electrical rate in the US. There are forecasts that will go to seven or 8% um in the next 5 to 10 years. It may not be that much, but it's directionally correct. China is preparing for that kind of growth. I mean, uh, people here know the statistics and scale. So obviously the topic here today is energy and AI, um, and concern about our generating capability and the and the health and resilience of our grid. At this point, is energy holding back the development of AI in the US, or do you foresee that happening in the short term? So I do a weekly call with the heads of all the AI companies, which is moderately shocking, and in this call we rank the problems, and because it's—you can't talk about private stuff in the companies, obviously, but you can talk about shared problems, and energy is the number one or number two problem; the other one is Washington—uh—and and their interrelation, of course. The complaint about Washington is that these companies need to make sure that nobody's trying to be helpful by preventing them to slow—by slowing them down incorrectly. So when you're inventing something this powerful, you—we seek oversight, but we don't want that oversight to turn into regulation. I'll give you an example: 10 years ago, Google decided to talk to—spent a lot of time with California on the self-driving cars at the time, and as a result, California banned them, right? And we we had this internal debate as to whether this was a good idea or not. We ultimately decided it was better to tell people what we're doing and be banned than the other way around, and then we unbanned it, right? We eventually said, "You guys are just wrong," and eventually the right thing happened. So it's an example of how it's tricky, um, but we don't have a choice; we have to—we collectively in the industry have to work with the governments. Google made the right decision then; we're making the right decision now. In terms of energy, the current belief is that we run out of power in the US in roughly four years. It's uh—and people come to those numbers by various ways, and they're all approximations, but it goes something like this: if you look at chip availability right now, uh H100s and their and their competitors are readily available. Everyone's excited about the B200, and the B200—a part of the reason is NVIDIA built this huge chip, which on most metrics uses less power; it's really an amazing achievement—and this huge chip puts out more than a kilowatt. Um, the packaged versions, which are two chips plus a essentially a GPU—CPU—in their case, it's an ARM chip—uh, that package needs liquid cooling, so that's causing the data centers to have to have to be upgraded, which is a non-trivial expense. But the important thing is when you do the math based on the B200 and B200A and you assume a cost-learning curve, you get huge improvements in energy efficiency per calculation. The problem is that if you look at what they want to do with it, they are 10 times more expensive for the thing that they're subsidizing. So a round number—not exact number—is a Google search is one; then the equivalent um essentially LLM search is a factor of 10 computationally, which translates into energy—that in—holding CPU constant. So so the long point here is that it's the—but you have a a this enormous benefit that the physicists are giving us—these larger chips—which mean that you want to be on the latest chips; the latest chips in aggregate use more power because you want—so more money; that's the tradeoff. So, given Washington, can we build the power capacity that we need for this industry in the time frame that we do? So how long from the time I asked for a transmission line in the United States does it take to get it permitted and approved? More than 10 years. I'd—yes, and the answer is 18. Uh, you know, I—at my age, I'm worried about being dead by then, uh, and I'm serious. So you all have more time to screw around with this problem. Uh, you can't run a flexible power grid if your opportunities for learning are every one—every 18 years. The learning cycles in my industry are weekly, right, or monthly, right? This continuous learning is how these things are done. If just to dump on—dump on on the US for a while, uh, I spent a lot of time in the Pentagon; the procurement cycles are so long that there's no possibility for iterative learning, and they only have a specialized model to do that, which is called OTAA, and OTAA is a very diff—is is uh specialized in various ways that are complicated. So because the government is not organized for a rapid change, the laws and the regulations make it very very difficult to move quickly. If you look at um—so let's go through the choices: new solar and new um wind are materially cheaper than new gas, and yet you have to put in gas peer plants for the two weeks in which the wind doesn't blow, and all of this—everybody here in the room—well, AI should build us a better battery. Um, AI will, in the form of better—better material science—how long does it take for a new—it to go from the lab in AI to a better battery? 5 years? 6 years? 7 years? It's not a—the learning cycles are just too slow, and the people are working hard on batteries; there's a lot of money there. Um, if you could build large—large storage batteries, uh, you eventually run out of the ingredients of those batteries; the command—the demands are so great. So the—speaking as—a lot of those ingredients come from China—well, a lot of the AGs come from wacky countries like Bolivia and various places in the in the Congo and so forth. Um, the important thing—of—is rare earth materials—China as well. I think the most important thing that we can do right now is to get on a war footing about producing more energy in the places where people are going to use them. There's plenty of energy; it's in the wrong places. You all understand that, speaking as—as your networking person—is what I've done my whole life. Why is this not a network switching problem that can be handled by computers? I used to—way back when, when I was building networks—I had to learn the programming of the modems, and I memorized the AT09—whatever the old-timers will remember this—this is completely forgotten now because you don't have to reprogram your modems; the system does it automatically. Why is that? Why is it not possible to produce—so if you look at the way the utilities are managed, they're managed to basically maintaining constant 60-cycle power, and they have an error bar—that problem should be solvable by centralized command of those—of those networks pretty well. I looked at the uh renewable portfolio standards, and here's the problem: if you assume the RPS is—maybe depending on what number is an assumption—is you make 20 or 30% of American planning—and current planning—that still leaves 70% in the traditional power plants, um, and we're running out of base load unless you're willing to build a lot more coal, which we're not willing to, and and more gas, which we are willing to. I don't see how you can get much above 50% in renewable, which means you're always going to have the intermittency—this is my opinion—the intermittency problem—between the two—and that problem has to get solved, and it gets—and what I would do—again, speaking to the government—is do a test—bit of that—demonstrate it and do it this year. I'm sure it can be done. So um, OpenAI is talking about taking an international approach and building—building data centers all around the world. Um, your reaction to that, and does that pose a security risk? So um, I'm involved in a lot of discussions with the Arab states: Saudi, uh, Qatar, UAE. I won't talk specifically about OpenAI, and they're interesting because they have lots of power, lots of sunshine, and they like to build things. We don't—we forgot how to build things; they build these things very very quickly; they have lots of empty space, you know, and so forth. And I'm in meetings with them in the next few—few months, and their plans are big: 50 billion, 70 billion, 100 billion. So that—that money is going to go in; they're going to build the data centers, and they also have a big commitment to renewable power, so they may be a pretty good test bed for how we should do it. And I mean that with—with um with praise of them. You have a lot of really smart people; they're spending a lot of money, and all the top US companies will be the beneficiaries of all that money as well. Isic. Now the question about where do you do your training and where do you do inference? I'll give you an example. Um, this—the the Arab world, at a minimum, will be the only training center for Europe, because the European energy prices are so high now, for all the reasons that you all know, um, that it just doesn't make sense to—you're training in Europe; you can—
I worked on this in France with Macron and so forth. If you call up EDF, which is their electric supplier, you can negotiate a special deal. The special deal is still more expensive, which is a good thing; it unfortunately, it's still more expensive than the training cost in the US or in the, or will be in the Arab world. So virtually all of the uh European models are trained elsewhere, usually in the United States, and that's not bad. Um, in China, there there're um I'll give you the example, they're about six or 700,000 GPU clusters, are better, probably a couple at Microsoft, probably a couple at Google. I'm these are my opinion; uh there's clearly one or more in Meta; uh there's clearly one Anthropic; um there's one um from not not BYU but um uh anyway, we'll come back with the WeChat company Tencent in Malaysia, and uh there's a couple of others being built. So I would watch where those are getting built to understand where the real power is going to be because if they're being built in the US and in China or in Malaysia, then that tells you where where the action will be from a national security perspective.
Um, one way for the National Security commit, and I've been in conversation with this, but let me just say my opinion, you have um the top five are the five eyes, and let's assume for purposes of argument that we trust France and Germany and and a few other countries, maybe Japan and Korea. Let's let's make a list like that, not not my I'm not the expert to make it, but let's say there's eight countries. There's probably another set of countries which we like a lot but are not quite as trusted; the first a good example would be India. We like India a lot, but they're not completely aligned with us for all the reasons that you know.
Um, I think the Arabs will come in at that level in the sense that um how secure do we think their work is, or is it possible that they've been infiltrated by the bad guys? You know, it's harder to know, hopefully not. Um, and one way to solve that problem is that the secure training loads go on in the five eyes, and the ones for everything else goes on anywhere. Um, and there's a whole separate question of Securities models, but that the security problems exist wherever you train; they have, you know, can they do cyber attacks, can they do biological attacks? It's not a it's not a training issue; that's a governance issue.
You spent some time talking about Renewables; um there was a fleeting reference to coal; um what about nuclear? You haven't really talked about the nuclear option. You know, Microsoft is talking about history; I'm sorry, the other nuclear option, not the one of all yesterday. We record the German decision to shut down their perfectly fine working fission plants is a terrible idea, and I think it's generally understood we should uh we should not be turning down; we should not have been um so the one in California was perfectly fine; it could have been kept long longer. Three Mile Island should be reopened; it should absolutely these these systems are quite mature and quite safe. In this year, in this year, how many people died of nuclear reactions? Basically zero. In this year, how many people will die die from coal production in the world? The number is close to half a half a million people because coal is really bad. To M, so I appreciate all the concerns, but this is a hungry world, and we're going to build it; we're going to build the energy one way or the other.
Secretary Granholm suggested in her remarks that perhaps some of the tech companies should be working in concert, um for instance to order modular nuclear plants; that this would improve the likelihood of success as well as so I'm going to put in my purchase order today. Okay, today, and for purposes of argument, the estimate is $2 billion; we all know it'll be 10, but I put in my order for two because that's how much money I convince the banks to give me. Who do I give it to? Westinghouse just pulled out of South Carolina. The best SMR, small reactors, are basically in Korea; they're not fully certified here, right? Um, I mean, I can go on; the the standardization what happened in nuclear is that, and I will get this in, there are people here who know this better than I, when they wrote the rules 30 years ago, the standard for safety of nuclear had to be absolute, not relative to background radiation, and that problem means that the systems are over-engineered right in a way that makes them very expensive. So I personally, I first I don't know who to give the purchase order to. Second, the typical approval right rate for new uh reactors in the United States is 11 to 13 years; that's considered fast. Um, again, there's no possibility of a learning cycle, and so I'm assuming that we'll build a few more Fusion plants, which is fine; the future will be Fusion. I'm involved with one of them; Billy and his team have a whole Fusion task force; syst been set up; it it feels like Fusion is available in a decade, but even that's not fast enough for this uh this problem now.
Um, Dr. Schmidt said he'd be happy to take some audience questions, and I'm sorry to to be so blunt, but I figured it's just more fun. Um, I mean, the fact of the matter, it's not like we haven't been having these conversations for a long time. We have a reason to fix this stuff, guys. We have enormous Demand on something which is a national security; it's a huge thing for America, right? If we get this right, that's why this meeting is so important; that's why what you're doing is so important, and and that's why I'm pushing so hard. Sorry, that's okay. Do we have some questions out there? I will try to see your hands. We have a couple of mic Runners; certainly, we have some; we have a couple right over here. Oh, here we have one; if we can get a mic over here, and please tell us if you wouldn't mind your name and who you're with, and if you could keep your questions crisp, that would be great. We'll get tomorrow.
Uh, good afternoon, Dr. Schmidt. Uh, Bill Bowman, I'm a fellow at State Department. Yes. Hi. Hello. Thank you for being here. Uh, you mentioned that training it's going to be in the airport for uh Europe; is that safe to say that then the data centers for imprints will be in Europe, and slow will they be closer to City centers, or could you talk about some of the data center um I'm I'm a big fan of fiber every, and I don't think there's that much latency between the Arab world and and Europe in terms of delivering real-time systems; there's so many other delays. I could be wrong. Um, and in general, when you use Google, you don't know where the server is, and as you travel around the world, you assume that it's in your country; it's seldom in the country that you're in; it's usually in a neighboring country. So when I think about this problem, I don't think of this as training versus inference; I think of this as the total energy cost and the electricity cost, and there's also a debate in the industry is what percentage will be training versus inference, and it goes something like this: the people who want to do training think that the mixture of expert models and the roles that they're doing every 18 months will essentially invent new systems faster than the inference can take in; other words, this the back end gets too fast, and so you shouldn't overweight inference. There's other people who think that that's not going to occur, and therefore the majority of the of the world load will go into inference, which will then be used to train better models. If you follow the reasoning, and we don't know the answer to that question, so my assumption again could be wrong is that you're going to see training and inference in the same location, partly because of reuse of Hardware, right? Especially if you buy all this stuff, it's incredibly expensive. A GPU is $50,000, fully burdened; a petabyte of storage is now about $50,000, again fully burdened. If you have 100,000 uh GPUs, you have five billion dollars, whatever it is, and then you have to depreciate that over because cycle is every two years. If you depreciate that every two years, it makes sense to me that you probably want to centralize it. The other reason is that because of speed of light issues, these systems are getting smaller and Tighter and Tighter uh because of memory use, and the way the algorithms work is they essentially take a very large amount of DRAM and they just go across it and back and forth and back and forth; that's called gradient descent, and um those those memory units need to be extremely close physically to the training.
We have another question right here. Oh, hi. Are you V Maradi and from Defense and Aerospace Report? Gene. Terrific job, sir. Great to see you. Terrific event. We do what we do because we do it, not because it makes sense, right? At almost every place we run into this, whether it's at the Pentagon and how they're thinking about big problems, whether it's Congress, whether it's in the energy grid and elsewhere in this highly competitive contested time, what what's it going to take to move the needle? We have folks like you who've been talking about this not just for one year but 10 years, 15 years, and yet we keep hitting the iceberg; we keep hitting the grey Rhino, whatever uh analogy you want to use; what's it going to take to actually move the needle in these things? We've had a lot of success talking to the Congress, shockingly, and because people understand energy in a sort of Base way, they understand what it does and why you need it, and when you show them the curves, they get it, right? The problem here is you don't have a single operating Authority; as you know, there are three grids: East Coast grid, West Coast grid, Texas. It looked like the Texas grid was the best, and they had various problems, including running out of power in a snowstorm. So again, it's just a mess. Uh, my general advice would be to try to get on the grid side; have the have have the the industry as a whole make a proposal to to go from 3% of energy to 8% energy in six years. What does it take, right? The sighting and the transmission lines, just show it. I did all this at Google, and uh we did it fairly quickly because the power was already there, so we would do is we would trade, you know, we would trade off Iowa and Ohio, um you know, for better terms; ultimately, we built in both, I might add. Um, and you trade them off; you get everybody focused on the competition; you get the extra lines; you put the data center literally right now next to these incredibly powerful high voltage Towers; you then have agreements over how much of that is renewable, often nuclear versus uh non-renewable, and you know, you work on all of that; it takes the whole team, which Google has, and I'm sure the other guys have as well. I read an article which I have no idea is true or not that said that Microsoft is going to build a data center uh larger than the sum of the Google data data centers in Wisconsin. Good luck on that, right? The amount of power and sighting, and what we would do is we would basically build bu all this land; we'd build one pad, put the data center in, and as the hardware was going into the First Data Center, we would build a second pad, so we always had a building ready, and that's what it takes. So so for me, the national plan would be identify what's needed and then clear areas where this stuff can be done with appropriate State Control agreement subsidies and so forth, and then just let it build, right? The the thing that is extraordinary now is how much money there is on the company side to do this; it's like infinite compared to what I what I was dealing with.
Uh, we have another question here. Uh, good after my name is Jamie Kar; I work on energy policy at the delegation of the European Union to United States; uh appreciate the conversation, um but I I throughout the day noticing that for the summit on AI and energy it's been unidirectional and talking about how much energy will we need and can we build for AI, but what about the other direction? Can you talk about how AI can be used to reduce the uh existing and future Demand on energy, not just in the data centers but across the power grid and deves we use in the way in which buildings and systems and individuals consume electricity in a way that it could potentially offset its own needs? Uh, haven't heard that discussion other than the brief mention you had about build a better battery. So so there's an old saying: I was at a dinner with Bill Gates and Andy Grove once, and they had a huge fight; this is years ago, uh and the saying was Grove giveth Gates taketh away, so the hard guys get better and they do a really good job, and the softer people just need more, and I have every reason to believe that in this phase of AI the same is true; it won't be Grove and Gates; it'll be some other people, but the fact of the matter is the demand is infinite, even assuming everything that you're implying in your question. So let's go through that first: we've already discussed better materials for batteries; that takes a while, but that's important. Um um you move to DC uh power lines for better better uh transmission loss control, things like that; you build those things; that takes a while. Um, in terms of delivery of power, you work on power line losses; you put the data centers next to the big power lines as I described earlier, and architecturally there huge improvements to be made in uh essentially cost Effectiveness per per transaction; all of that will be swamped by the enormous needs of this new technology, and because it's a universal technology and because it's the arrival of an alien intelligence, right? Something we don't understand; we may we may make mistakes with respect to how it's used, but I can assure you that we're not going to get there through conservation; that's the key point, and the economics will drive it anyway. No large company wants to have a huge power bill; most of the people I've talked with say power bill is become a very large component of their expenses.
Do you think we can meet AI's energy needs without totally blowing out climate goals? It it depends on well, so the first place is there are plenty; the US uses less power now than we did uh and generates less CO2 than we did uh of the bad kind over the last five or 10 years, so the industry and the regulatory structures are working to reduce our percentage BAS basis and so forth, so it's really kind of a a race between improvements in, for example, gasoline emissions, electric cars, all the things that you all know; again, that is getting much better; good job; and then this over thing. My own opinion is that we're not going to hit the climate goals anyway because we're not organized to do it; we're not, and the way to do it is with the things that we're talking about now, and yes the yes the the needs in this area will be problem, but I'd rather bet on AI solving the problem than constraining it and having the problem. If you see my plan, we're at a time, but top three action steps that need to be taken now in your opinion? Well, first place, the industry is working incredibly hard on this, and what the industry needs from you all is more power that is predictable that is base load or Bas load equivalent as soon as they can; they need more places to site; they need more ways of getting things built, and they get need that need have need more things to get them connected. In return for your good work, I promise you that we will make your systems and in particular your generation systems at least 15% better, and that's a lot of money for a utility. Why do I say that? Because at DeepMind when we did the same thing for our beautifully crafted handcrafted data centers with the best engineers in the world, according to us, this nasty little system that we built beat our best Engineers by 15% in total Energy Efficiency, and there were cases where it was a 50%, and it's because the AI systems see things that humans don't: the changes in load predictability and so forth. It's a reinforcement learning problem, and it can be done efficiently. I think that we can do this, and I feel very strongly, and and let me just finish by saying GD and for you as always, thank you for leading in this area. Illy, thank you for doing this, and thank you all for coming. I cannot think of a better project to work on this year. Thanks again. Got this note. Thank you so much.