1 00:00:02,160 --> 00:00:04,077 I honestly believe that the AI 2 00:00:04,160 --> 00:00:09,437 revolution is underhyped. Now, why is this all important? [music] 3 00:00:09,520 --> 00:00:12,637 >> Eric Schmidt is here. He's the former Google executive chairman and CEO. 4 00:00:12,720 --> 00:00:15,637 >> These agents are going to be really powerful, and they'll start to work 5 00:00:15,720 --> 00:00:19,037 [music] together. We're soon going to be able to have computers running on their 6 00:00:19,120 --> 00:00:22,877 own, deciding what they want to [music] do. Now we have the arrival of a new 7 00:00:22,960 --> 00:00:29,037 nonhuman intelligence which is likely to have better reasoning skills than humans 8 00:00:29,120 --> 00:00:31,117 can have. >> So if [music] you were emperor of the 9 00:00:31,200 --> 00:00:35,117 world for 1 hour, the most important thing I do is make [music] sure that the 10 00:00:35,200 --> 00:00:36,877 west wins. 11 00:00:36,960 --> 00:00:40,637 >> Ladies and gentlemen, please welcome 12 00:00:40,720 --> 00:00:44,200 Eric Schmidt. 13 00:00:44,960 --> 00:00:47,220 >> Hi. [music] 14 00:00:47,840 --> 00:00:51,007 Hi. Looking good. Good. Thank you. 15 00:00:51,090 --> 00:00:52,477 [applause] >> Good to see you. 16 00:00:52,560 --> 00:00:57,277 >> Good to see you. >> You're looking, Eric. You're looking very nice. 17 00:00:57,360 --> 00:01:00,557 >> Oh my god. David, good to see you. It's like >> David Sax is here as well. 18 00:01:00,640 --> 00:01:04,877 >> It's like a It's like a reunion of all of our former companies. David, why did 19 00:01:04,960 --> 00:01:07,757 you quit after all? >> My old [clears throat] 20 00:01:07,840 --> 00:01:12,237 >> my old boss. [laughter] It was >> What was it like working with Young 21 00:01:12,320 --> 00:01:16,397 Freedberg? Take us back. >> Can I Can I tell a story? We have to 22 00:01:16,480 --> 00:01:19,277 come down to Orange County and they're like, "Hey, we're going to take the 23 00:01:19,360 --> 00:01:23,037 plane and it was Eric's plane." We get on the plane and then he goes up and 24 00:01:23,120 --> 00:01:24,957 flies the plane. I'm in the back of the plane by myself. 25 00:01:25,040 --> 00:01:25,917 >> Was it King Air? What is it? 26 00:01:26,000 --> 00:01:28,717 >> I'm like, "The CEO of Google's flying me down to Orange County." It was 27 00:01:28,800 --> 00:01:31,837 incredible. That was my first time actually hanging out with Eric. 28 00:01:31,920 --> 00:01:34,077 >> It was It was my Gulfream. >> That's right. 29 00:01:34,160 --> 00:01:37,837 >> Um, he was way too smart. >> Way too smart. 30 00:01:37,920 --> 00:01:41,357 >> Way too smart. >> Was he focused? Did he contribute? Did he move the needle? 31 00:01:41,440 --> 00:01:46,477 >> But he was very smart. [laughter] Okay, that's kind of our consensus off the pod 32 00:01:46,560 --> 00:01:49,117 as well. >> Look, you guys know this guy well. He's 33 00:01:49,200 --> 00:01:51,917 really that smart. So, he taught me more 34 00:01:52,000 --> 00:01:56,157 stuff than most of any of the employees at Google. And then you left. 35 00:01:56,240 --> 00:01:57,677 >> Well, tell us what you've been doing. 36 00:01:57,760 --> 00:02:00,397 So, um >> No, no, before that, I got to ask you 37 00:02:00,480 --> 00:02:04,157 this question. There was a recently deleted video from Stanford. 38 00:02:04,240 --> 00:02:07,277 >> Oh, no. >> You had a moment of clarity where you 39 00:02:07,360 --> 00:02:11,277 said, "Hey, you know, like at Google, people are like too much work life 40 00:02:11,360 --> 00:02:15,357 balance. They need to commit. They need to work harder. We had Sergey at the 41 00:02:15,440 --> 00:02:21,720 last event. He's going back to work. So Sergey got the message. 42 00:02:22,000 --> 00:02:25,987 Predicting Sergey's behavior is something I can fail at. I tried for 43 00:02:26,070 --> 00:02:28,877 [laughter] 20 years. Um I am not in 44 00:02:28,960 --> 00:02:33,117 favor of um essentially working at home. 45 00:02:33,200 --> 00:02:35,757 I And the reason I mean many of you guys 46 00:02:35,840 --> 00:02:37,837 all work at home to some degree, but 47 00:02:37,920 --> 00:02:40,157 your careers are already established. 48 00:02:40,240 --> 00:02:44,717 But think about a 20-some who has to learn how the world works and you know 49 00:02:44,800 --> 00:02:47,517 they they come out of Berkeley or Dartmouth and they're very well 50 00:02:47,600 --> 00:02:50,157 educated. When I think about how much I 51 00:02:50,240 --> 00:02:54,557 learned when I was at Sun just listening to these elder people who were 5 or 10 52 00:02:54,640 --> 00:02:57,037 years older than I was argue with each 53 00:02:57,120 --> 00:03:00,317 other in person. I don't how do you 54 00:03:00,400 --> 00:03:04,957 recreate that in this new thing? And and I'm in favor of work life balance and 55 00:03:05,040 --> 00:03:06,317 that's why people work for the 56 00:03:06,400 --> 00:03:11,117 government. Um sorry. Um, strays. 57 00:03:11,200 --> 00:03:14,157 >> Sorry, sorry, sorry. Um, if you're going 58 00:03:14,240 --> 00:03:17,677 to be in tech and you're going to win, you're going to have to make some 59 00:03:17,760 --> 00:03:19,917 trade-offs. And you're remember, we're 60 00:03:20,000 --> 00:03:23,197 up against the Chinese. The Chinese work 61 00:03:23,280 --> 00:03:25,677 life balance consists of 996, which is 62 00:03:25,760 --> 00:03:28,717 9:00 a.m. to 9:00 p.m. 6 days a week. By 63 00:03:28,800 --> 00:03:33,117 the way, the Chinese have clarified that this is illegal. However, they all do 64 00:03:33,200 --> 00:03:35,757 it. >> That's who you're competing against. 65 00:03:35,840 --> 00:03:39,277 >> I brought I brought everybody back to office. It's so much better. So, it 66 00:03:39,360 --> 00:03:42,637 let's just pick up on that theme. So, >> you don't need to defend the government. 67 00:03:42,720 --> 00:03:45,997 >> No, no. Believe me, I don't I don't see the need to. I'm an unpaid part-time 68 00:03:46,080 --> 00:03:48,877 adviser to the government. So, uh but uh 69 00:03:48,960 --> 00:03:51,437 but we are in this high-tech competition 70 00:03:51,520 --> 00:03:53,597 with China. They obviously care about 71 00:03:53,680 --> 00:03:56,157 AI, too. They're trying to race ahead. 72 00:03:56,240 --> 00:04:00,637 How do you uh I I understand that you recently ma made a trip there. How do 73 00:04:00,720 --> 00:04:04,077 you um handicap this this this competition? Well, you and I just talked 74 00:04:04,160 --> 00:04:07,677 about this as part of your as your incredibly important 75 00:04:07,760 --> 00:04:10,397 >> work in the White House. Um, I had 76 00:04:10,480 --> 00:04:12,237 thought that China and the United States 77 00:04:12,320 --> 00:04:15,837 were competing at the peer level in AI 78 00:04:15,920 --> 00:04:19,437 and that the good work that you have done and your predecessors did to 79 00:04:19,520 --> 00:04:21,677 restrict chips were slowing them down. 80 00:04:21,760 --> 00:04:23,437 They're really doing something more 81 00:04:23,520 --> 00:04:25,437 different than I thought. They're not 82 00:04:25,520 --> 00:04:29,517 pursuing crazy AGI strategies partly 83 00:04:29,600 --> 00:04:33,037 because the hardware limitations that you've put in place, but partly because 84 00:04:33,120 --> 00:04:35,357 the depth of their capital markets don't 85 00:04:35,440 --> 00:04:38,157 exist. They can't raise based on a wing 86 00:04:38,240 --> 00:04:40,957 and a prayer $100 million or maybe an 87 00:04:41,040 --> 00:04:42,877 equivalent to to build the data centers. 88 00:04:42,960 --> 00:04:45,837 They just can't do it. And so the result 89 00:04:45,920 --> 00:04:48,317 is they're very focused on taking AI and 90 00:04:48,400 --> 00:04:53,037 applying it to everything. And so the concern I have is that while we're 91 00:04:53,120 --> 00:04:56,317 pursuing AGI, which is incredibly interesting and we should talk about and 92 00:04:56,400 --> 00:05:00,237 all of us will be affected by this, we better also be competing with the 93 00:05:00,320 --> 00:05:04,397 Chinese in day-to-day stuff. Consumer apps, this is something you understand 94 00:05:04,480 --> 00:05:08,557 very well, Chimoth, uh, consumer apps, uh, robots and so forth and so on. I saw 95 00:05:08,640 --> 00:05:11,037 all the the Shanghai robotics companies 96 00:05:11,120 --> 00:05:13,837 and these guys are attempting to do in 97 00:05:13,920 --> 00:05:18,397 robots what they've successfully done with electric vehicles, right? and 98 00:05:18,480 --> 00:05:22,237 they're re they their work ethic is incredible. They're well funded. It's 99 00:05:22,320 --> 00:05:24,957 not the crazy valuations that we have in 100 00:05:25,040 --> 00:05:29,437 America. They can't raise the capital, but they can win across that. The other 101 00:05:29,520 --> 00:05:32,477 thing the Chinese are doing, and I want to emphasize this is a major 102 00:05:32,560 --> 00:05:36,797 geopolitical issue, is that my own background is open source. In the 103 00:05:36,880 --> 00:05:38,477 audience, you all know open source means 104 00:05:38,560 --> 00:05:41,037 open code. Open weight open weights 105 00:05:41,120 --> 00:05:45,677 means open training data. China is competing with open weights and 106 00:05:45,760 --> 00:05:49,757 open training data. And the US is largely and majority focused on closed 107 00:05:49,840 --> 00:05:54,077 weights, closed data. That means that the majority of the world, think of it 108 00:05:54,160 --> 00:05:57,837 as the belt and road initiative, are going to use Chinese models and not 109 00:05:57,920 --> 00:06:02,717 American models. Now, I happen to think the West and democracies are correct. 110 00:06:02,800 --> 00:06:05,997 And I'd much rather have the proliferation of large language models 111 00:06:06,080 --> 00:06:07,357 and that learning be done based on 112 00:06:07,440 --> 00:06:09,837 Western values. Eric, we had a a major 113 00:06:09,920 --> 00:06:13,357 open- source initiative um with Meta, 114 00:06:13,440 --> 00:06:17,917 you know, incredible balance sheet, tremendous technical firepower, but they 115 00:06:18,000 --> 00:06:21,757 seem to have misexecuted and now are taking a step back and reformulating 116 00:06:21,840 --> 00:06:24,797 something to your point that looks a little bit more closed source. 117 00:06:24,880 --> 00:06:28,957 >> It's not clear. You know, Alex Wang's a good friend. Uh he's come in, he's taken 118 00:06:29,040 --> 00:06:33,037 over. He's obviously incredibly uh incredibly capable. I would not hold 119 00:06:33,120 --> 00:06:35,357 keep um I I would not say that they're 120 00:06:35,440 --> 00:06:37,517 going fully closed. And I think also 121 00:06:37,600 --> 00:06:40,237 they got screwed up because the deepseek 122 00:06:40,320 --> 00:06:44,397 people uh R1 did such a good job, right? 123 00:06:44,480 --> 00:06:48,717 If you look at the reasoning model in deepseek and in particular their ability 124 00:06:48,800 --> 00:06:51,757 to do reinforcement learning forward and back, forward and back and forward and 125 00:06:51,840 --> 00:06:54,477 back, this is a major achievement and it 126 00:06:54,560 --> 00:06:59,277 appears that they're doing it with less precision than numeric precision than 127 00:06:59,360 --> 00:07:01,357 the American models. As a bit of 128 00:07:01,440 --> 00:07:03,517 technical things, uh there's something 129 00:07:03,600 --> 00:07:07,677 called FP64, FP32, FP16. The American 130 00:07:07,760 --> 00:07:11,037 models are typically using 16 bit precision for their training. The 131 00:07:11,120 --> 00:07:14,077 Chinese are pushing eight and now even four. 132 00:07:14,160 --> 00:07:19,517 >> Is there is there something that um the American you know bigger companies need 133 00:07:19,600 --> 00:07:22,317 to be doing in open source so that we can actually combat this? 134 00:07:22,400 --> 00:07:25,197 >> Well, a number of the large companies have said that they want to be leaders 135 00:07:25,280 --> 00:07:27,597 in open source as well. Um Sam Alman 136 00:07:27,680 --> 00:07:29,837 indicated that the smallest version of 137 00:07:29,920 --> 00:07:32,717 the 03 model would be released I believe 138 00:07:32,800 --> 00:07:35,597 open weights and they have done so and 139 00:07:35,680 --> 00:07:38,877 he told me anyway that this model is 140 00:07:38,960 --> 00:07:41,677 much smaller than 10 the 26. It's much 141 00:07:41,760 --> 00:07:43,837 easier to train and it will fit or or 142 00:07:43,920 --> 00:07:47,597 can fit on your phone. So one path is to 143 00:07:47,680 --> 00:07:51,917 say that we'll have these supercomputers doing AGI which will always be 144 00:07:52,000 --> 00:07:55,677 incredibly expensive and so forth. But we also have to watch to make sure that 145 00:07:55,760 --> 00:07:57,837 the proliferation of these models for 146 00:07:57,920 --> 00:08:00,317 handheld devices is under American 147 00:08:00,400 --> 00:08:03,437 control whether it's open AAI or Meta or 148 00:08:03,520 --> 00:08:06,477 Gemini or what have you. recently you uh 149 00:08:06,560 --> 00:08:10,877 took over uh Relativity Space and I think for for the people that don't know 150 00:08:10,960 --> 00:08:15,757 this is a business that effectively whose ambition is to compete with SpaceX. 151 00:08:15,840 --> 00:08:18,957 >> I think you were the first investor or the earliest investor in it. 152 00:08:19,040 --> 00:08:21,437 >> I was >> I'm sorry. >> It's okay. [laughter] 153 00:08:21,520 --> 00:08:24,397 >> Lost some money in the first round. What happened? Did you get crammed down? 154 00:08:24,480 --> 00:08:26,157 >> No, no, no. I mean, >> all of us did. 155 00:08:26,240 --> 00:08:30,477 >> All of us. I mean, look, >> I mean, I've been very happily uh you 156 00:08:30,560 --> 00:08:34,477 know, an investor in SpaceX and Swarm and Starlink. relativity was 157 00:08:34,560 --> 00:08:37,597 >> and by the way swarm is a big deal created swarm was a >> so thank you 158 00:08:37,680 --> 00:08:41,197 >> yeah swarm has been a really great success for for them and I think for the 159 00:08:41,280 --> 00:08:43,517 world um but what I was going to ask you 160 00:08:43,600 --> 00:08:48,557 is walk us through the evolution of the space market why you decided of all the 161 00:08:48,640 --> 00:08:53,037 companies you have the capital base to kind of put your money anywhere why did 162 00:08:53,120 --> 00:08:56,157 you pick that why did you pick that business why now 163 00:08:56,240 --> 00:08:58,157 >> rockets are really cool and they're 164 00:08:58,240 --> 00:09:01,117 really hard I had I'm as you know I'm a 165 00:09:01,200 --> 00:09:03,437 pilot and I know lots about jets And I 166 00:09:03,520 --> 00:09:05,357 had assumed that rockets were as mature 167 00:09:05,440 --> 00:09:08,397 as jet engines. They're not. It is an 168 00:09:08,480 --> 00:09:13,437 art and a science. These things are very hard to do. The amounts of power, I 169 00:09:13,520 --> 00:09:18,157 mean, in our case, the rocket is 4 million pounds of thrust. Um, you have 170 00:09:18,240 --> 00:09:21,917 to hold the thing down to test it and you can't even hold it with metal 171 00:09:22,000 --> 00:09:24,397 things. You have to have other things to hold it down as well. There's so much 172 00:09:24,480 --> 00:09:28,237 force otherwise it will take off. Um, another interesting thing about rockets 173 00:09:28,320 --> 00:09:30,717 is that a rough number is that 2% of the 174 00:09:30,800 --> 00:09:33,437 weight of the rocket is the payload, 18% 175 00:09:33,520 --> 00:09:38,397 is roughly the rocket, and 80% is the propellant. And my reaction as a new 176 00:09:38,480 --> 00:09:42,397 person is, you're telling me you can't do any better. And the physicists say 177 00:09:42,480 --> 00:09:46,317 after 60 years of physics, that's the best we can do to get out of the 178 00:09:46,400 --> 00:09:50,477 gravitation of of the Earth. And so, I think rockets are interesting and 179 00:09:50,560 --> 00:09:54,157 they're challenging. There's always an opportunity for competition. um in 180 00:09:54,240 --> 00:09:56,157 relativity space's area, it's 181 00:09:56,240 --> 00:09:59,277 essentially a LEO com a LEO competitor. 182 00:09:59,360 --> 00:10:03,197 So low Earth orbit satellites, that sort of thing. The order book is full. We 183 00:10:03,280 --> 00:10:08,237 just have to launch the rocket and and this entry into space happened and I'm 184 00:10:08,320 --> 00:10:11,597 not sure how well known this is. So you can go as far as you want to go into 185 00:10:11,680 --> 00:10:15,837 this, but you've done a lot as well in next generation warfare as well. Do you 186 00:10:15,920 --> 00:10:20,477 want us just to talk about that and how you ended up there and what role that 187 00:10:20,560 --> 00:10:21,917 plays and just give us a landscape? 188 00:10:22,000 --> 00:10:24,797 Maybe David asked about the China question, but it's they're all kind of 189 00:10:24,880 --> 00:10:27,117 almost interrelated. >> Well, first place, I'm a software 190 00:10:27,200 --> 00:10:30,397 person, not a hardware person. I I explain to people that hardware people 191 00:10:30,480 --> 00:10:33,917 go to different schools than software people. Um, and they think slightly 192 00:10:34,000 --> 00:10:37,757 differently. So, I'm always at a limitation in these new industries. Um, 193 00:10:37,840 --> 00:10:41,037 I had worked for the Secretary of Defense and have a top secret clearance 194 00:10:41,120 --> 00:10:43,517 and all that. I was given a medal, etc., 195 00:10:43,600 --> 00:10:48,397 uh, for trying to help the Pentagon reorganize itself. And when the Ukraine 196 00:10:48,480 --> 00:10:50,477 war started, I was watching and I 197 00:10:50,560 --> 00:10:52,157 thought, well, here's an opportunity to 198 00:10:52,240 --> 00:10:55,917 see a country that has no navy u and no 199 00:10:56,000 --> 00:10:58,157 uh and no air force, how they do this 200 00:10:58,240 --> 00:11:00,957 with automation. And indeed, it has been 201 00:11:01,040 --> 00:11:03,277 a spectacular success as a matter of 202 00:11:03,360 --> 00:11:06,797 innovation. Um and outnumbered 3 to one 203 00:11:06,880 --> 00:11:10,397 with huge differences in um kinetic 204 00:11:10,480 --> 00:11:13,037 strength, weapons, mobilization, and so 205 00:11:13,120 --> 00:11:17,437 forth and so on. Ukraine has held on really quite well. Um and what's 206 00:11:17,520 --> 00:11:20,877 happening now is you're seeing essentially the birth of a completely 207 00:11:20,960 --> 00:11:22,877 new military national security 208 00:11:22,960 --> 00:11:28,077 structure. Um one way to think about it is that we all u so first place and I've 209 00:11:28,160 --> 00:11:32,397 I've seen it live and I will tell you that real war is much worse than the 210 00:11:32,480 --> 00:11:36,237 worst movies you have ever seen about war. And that's all I'll say. It's 211 00:11:36,320 --> 00:11:37,837 really horrific and it's to be avoided 212 00:11:37,920 --> 00:11:41,197 at all cost. Um and then right for 213 00:11:41,280 --> 00:11:45,437 obvious reasons and the the 214 00:11:45,520 --> 00:11:49,837 and I love all these people say well you know the wararm mongering talk be 215 00:11:49,920 --> 00:11:54,157 careful what you wish for because the other side gets a vote. When I started 216 00:11:54,240 --> 00:11:55,997 working and and trying to understand 217 00:11:56,080 --> 00:11:58,797 what Ukraine was doing uh Russia was 218 00:11:58,880 --> 00:12:03,917 pushed back and they've come back with a very very strong second and third round. 219 00:12:04,000 --> 00:12:05,677 So the enemy gets a vote in this 220 00:12:05,760 --> 00:12:09,117 situation. uh but to but to go on u the 221 00:12:09,200 --> 00:12:12,237 rough way in which war will evolve is 222 00:12:12,320 --> 00:12:14,957 first things will have to be very very 223 00:12:15,040 --> 00:12:19,917 mobile and very much not in fixed places. This takes out most of the 224 00:12:20,000 --> 00:12:21,917 military infrastructure that exists in 225 00:12:22,000 --> 00:12:25,517 the world. Um, things like tanks, um, of 226 00:12:25,600 --> 00:12:28,957 which we're now building a whole bunch more, even stronger tanks here in 227 00:12:29,040 --> 00:12:31,117 America, don't make any sense in a world 228 00:12:31,200 --> 00:12:34,157 where a 2 kilogram payload from a a 229 00:12:34,240 --> 00:12:36,717 well-armed drone can destroy the tank. 230 00:12:36,800 --> 00:12:38,397 It's called the kill ratio. And that 231 00:12:38,480 --> 00:12:42,157 drone costs retail $5,000, $4,000. The 232 00:12:42,240 --> 00:12:46,637 tank, the American tank costs $30 million. You can see the you can send an 233 00:12:46,720 --> 00:12:48,237 awful lot of those drones to destroy 234 00:12:48,320 --> 00:12:51,437 those tanks. Um the likely evolution 235 00:12:51,520 --> 00:12:55,997 goes something like this. So first people learn that drones are like rifles 236 00:12:56,080 --> 00:12:58,157 and like artillery. [snorts] So it's 237 00:12:58,240 --> 00:13:03,757 more efficient to use drones now than to use mortars, grenades, artillery. That's 238 00:13:03,840 --> 00:13:07,677 clear. If you just look at the economics, economics in terms of cost or 239 00:13:07,760 --> 00:13:10,557 effectiveness as it's called. Um the 240 00:13:10,640 --> 00:13:15,197 next thing that happens is that both sides develop drone capabilities which 241 00:13:15,280 --> 00:13:17,277 is what you're seeing now and each then 242 00:13:17,360 --> 00:13:19,917 becomes a war of drone against drone. So 243 00:13:20,000 --> 00:13:24,477 you have drone against anti- drone. And so then the shift moves to how do you 244 00:13:24,560 --> 00:13:28,637 detect the enemy drone and how do you destroy it before it destroys you. So 245 00:13:28,720 --> 00:13:30,797 the doctrine ultimately is the drones 246 00:13:30,880 --> 00:13:34,157 are forward and the people are behind. 247 00:13:34,240 --> 00:13:38,477 And I've seen operations in for example sitting in Kev where the Ukrainians are 248 00:13:38,560 --> 00:13:40,477 commanding things over Starlink I might 249 00:13:40,560 --> 00:13:43,437 add um in the distance in the distant 250 00:13:43,520 --> 00:13:47,677 war and they're very very effective. So we've solved the latency problems, we've 251 00:13:47,760 --> 00:13:49,197 solved the timing problems and so forth 252 00:13:49,280 --> 00:13:52,637 in that area. The ultimate state is very 253 00:13:52,720 --> 00:13:56,797 interesting and I don't think anyone has foreseen this. If you go back to our 254 00:13:56,880 --> 00:14:01,277 conversation about RL and planning, which is what you're seeing with AI, 255 00:14:01,360 --> 00:14:05,117 let's say that that we're on one side and we have a million drones and there's 256 00:14:05,200 --> 00:14:09,437 another side over here that has another million drones. Each side will use 257 00:14:09,520 --> 00:14:11,357 reinforcement learning AI strategies to 258 00:14:11,440 --> 00:14:14,637 do battle plans, but neither side can 259 00:14:14,720 --> 00:14:19,197 figure out what the other side's battle plan is. And therefore, the deterrence 260 00:14:19,280 --> 00:14:23,197 against attacking each other will be very high. Today, the way military 261 00:14:23,280 --> 00:14:26,237 planners operate is that they count weapons. They say, "Well, you have this 262 00:14:26,320 --> 00:14:30,077 many and I have this many and you can do this kind of a maneuver and so forth." 263 00:14:30,160 --> 00:14:33,917 But in an AI world where you're doing reinforcement learning, you can't count 264 00:14:34,000 --> 00:14:37,757 what the other side is planning. You can't see it. You don't know it. And I 265 00:14:37,840 --> 00:14:41,517 believe that that will deter what I view as one of the most horrendous things 266 00:14:41,600 --> 00:14:43,357 ever done by humans, which is war. 267 00:14:43,440 --> 00:14:47,197 Because unless there's a perfect balance between either side, there will be some 268 00:14:47,280 --> 00:14:50,797 mutual destruction of the drone supply like there would be with any artillery 269 00:14:50,880 --> 00:14:55,117 stock in traditional warfare. And whoever's left ends up winning like 270 00:14:55,200 --> 00:14:57,677 they're just >> Well, it's very important to understand 271 00:14:57,760 --> 00:15:00,797 that there's no winners in war. Um, by 272 00:15:00,880 --> 00:15:04,717 the time you had a drone battle of the scale I'm describing, the entire 273 00:15:04,800 --> 00:15:07,677 infrastructure of your side will be destroyed. The entire infrastructure of 274 00:15:07,760 --> 00:15:11,837 the other side will be destroyed. These are lose-lose scenarios. Isn't there 275 00:15:11,920 --> 00:15:16,317 like an an equilibrium though that that can also create where there because of 276 00:15:16,400 --> 00:15:19,757 that mutually assured destruction it there's a det or is that 277 00:15:19,840 --> 00:15:23,277 >> well I'm arguing that it's it's not a deterrence >> right 278 00:15:23,360 --> 00:15:25,677 >> that as deterrence can be understood as 279 00:15:25,760 --> 00:15:28,157 I want to hit you which I don't but I 280 00:15:28,240 --> 00:15:32,797 want to hit you so much but that if I do that the penalty is the penalty is 281 00:15:32,880 --> 00:15:34,957 greater than the value of me hitting you >> right 282 00:15:35,040 --> 00:15:37,997 >> and that's how deter that's how deter >> but that seems like an ad that that 283 00:15:38,080 --> 00:15:41,037 seems like a great um advantage antage 284 00:15:41,120 --> 00:15:46,237 and upside of this move to sort of drones and automation that we don't have today. 285 00:15:46,320 --> 00:15:49,757 >> Well, I there are many advantages to moving to drones and automation. One, 286 00:15:49,840 --> 00:15:51,917 they're much much cheaper, right? 287 00:15:52,000 --> 00:15:53,597 They're much much cheaper. Yeah. 288 00:15:53,680 --> 00:15:57,997 >> And two, and two, you can stockpile algorithms. You can essentially learn 289 00:15:58,080 --> 00:16:02,237 and learn and learn. And remember, you can also build training data, right, 290 00:16:02,320 --> 00:16:05,597 that's synthetic, so you can be even better than the others. The final 291 00:16:05,680 --> 00:16:08,397 question I've been asked by our military 292 00:16:08,480 --> 00:16:11,117 is what's the role of the of a 293 00:16:11,200 --> 00:16:14,797 traditional land army. And I wish I 294 00:16:14,880 --> 00:16:17,437 could say that all of these human 295 00:16:17,520 --> 00:16:19,597 behaviors can occur without humans being 296 00:16:19,680 --> 00:16:22,797 at at risk. I don't think so. I think 297 00:16:22,880 --> 00:16:26,237 that the way um robot war essentially 298 00:16:26,320 --> 00:16:30,717 drone war will occur is there will be these destructive waves, but eventually 299 00:16:30,800 --> 00:16:33,757 humans are going to have to cross a line. they're going to have to 300 00:16:33,840 --> 00:16:37,277 >> after we've depleted them. So, you're investing in this drone technology and 301 00:16:37,360 --> 00:16:39,037 then do you think 302 00:16:39,120 --> 00:16:41,437 >> Optimus and humanoid robots are the 303 00:16:41,520 --> 00:16:45,517 next, you know, volley in this um new 304 00:16:45,600 --> 00:16:48,077 warfare. It's going to be a long time 305 00:16:48,160 --> 00:16:50,717 before humanoid robots we which we see 306 00:16:50,800 --> 00:16:53,277 in the movies all day, right? Be a very 307 00:16:53,360 --> 00:16:55,837 long time before we see that. Uh what 308 00:16:55,920 --> 00:17:00,637 you're going to see is very very fast mobility solutions, right? airly 309 00:17:00,720 --> 00:17:05,037 airbased solutions and also >> hypersonics >> hypersonics also things underwater 310 00:17:05,120 --> 00:17:06,797 there's a lot of that going on it's a 311 00:17:06,880 --> 00:17:09,837 different domain um if you look at the 312 00:17:09,920 --> 00:17:12,237 um the Muro and some other boats that 313 00:17:12,320 --> 00:17:14,077 the Ukrainians used they have 314 00:17:14,160 --> 00:17:17,837 essentially used USVs to destroy the uh 315 00:17:17,920 --> 00:17:20,637 Russian fleet in the Black Sea this was 316 00:17:20,720 --> 00:17:22,797 crucial for them because they needed to 317 00:17:22,880 --> 00:17:25,597 be able to export the grain from Odessa 318 00:17:25,680 --> 00:17:30,877 around and it's like 6% or 10% of their economy It's a very big deal and they 319 00:17:30,960 --> 00:17:33,517 did that with drones. >> Eric, it seems like there's this 320 00:17:33,600 --> 00:17:36,397 overarching worldview that you have, 321 00:17:36,480 --> 00:17:40,077 meaning you have this view on AI. 322 00:17:40,160 --> 00:17:44,637 There's all the stuff you're doing now in drones, in warfare, in rocketry. It 323 00:17:44,720 --> 00:17:48,157 all converges quite honestly because in the in the next 5 or 10 years, these 324 00:17:48,240 --> 00:17:53,517 things will all come to pass. What is the like how do you view the world? Like 325 00:17:53,600 --> 00:17:55,117 what is the role of America? What is 326 00:17:55,200 --> 00:17:57,517 your role as a as a capitalist, as a 327 00:17:57,600 --> 00:18:00,957 technologist, as like a statesman? 328 00:18:01,040 --> 00:18:06,577 >> I want America to win, right? 329 00:18:06,660 --> 00:18:07,357 [applause] 330 00:18:07,440 --> 00:18:11,597 Uh I am here because the American dream, 331 00:18:11,680 --> 00:18:15,837 the people who invested in me, in my case, Berkeley and so forth, people took 332 00:18:15,920 --> 00:18:17,837 a chance on me. I want the next 333 00:18:17,920 --> 00:18:20,397 generation to have that. I also want you 334 00:18:20,480 --> 00:18:23,117 all to remember I was just in in uh as 335 00:18:23,200 --> 00:18:27,677 part of the World War II surrender ceremony in in Honolulu and they talked 336 00:18:27,760 --> 00:18:29,997 about fighting tyranny, right? We forget 337 00:18:30,080 --> 00:18:32,477 that our ancestors or great-grandparents 338 00:18:32,560 --> 00:18:35,277 or whatever fought the Great War to keep 339 00:18:35,360 --> 00:18:37,997 liberalism and democracy alive. I want 340 00:18:38,080 --> 00:18:40,557 us to do that. How do we do that as 341 00:18:40,640 --> 00:18:43,277 Americans? We use our strengths. What 342 00:18:43,360 --> 00:18:46,237 are our strengths? We're chaotic, 343 00:18:46,320 --> 00:18:48,877 confusing, loud, you know, but we're 344 00:18:48,960 --> 00:18:51,837 clever. Uh we allocate capital smartly. 345 00:18:51,920 --> 00:18:56,157 We have very deep financial markets. We have this enormous industrial base of 346 00:18:56,240 --> 00:19:00,077 universities and entrepreneurs which are represented here. We should celebrate 347 00:19:00,160 --> 00:19:04,637 this. We should stoke it. We should make it go faster and faster. I spent lots of 348 00:19:04,720 --> 00:19:08,797 time in Europe because of the Ukraine stuff. They are so envious of us. When 349 00:19:08,880 --> 00:19:11,517 you're in Asia, they are envious of us. 350 00:19:11,600 --> 00:19:13,837 Don't screw it up, guys. That's what I 351 00:19:13,920 --> 00:19:14,717 want to work on. 352 00:19:14,800 --> 00:19:19,037 >> Can can I can I just ask you outside of 353 00:19:19,120 --> 00:19:22,797 this external conflict? We had um a 354 00:19:22,880 --> 00:19:26,637 conversation with Alex Karp today and we actually had Tucker Carlson here 355 00:19:26,720 --> 00:19:31,277 yesterday and some of the dialogue was around the I I don't know if the right 356 00:19:31,360 --> 00:19:34,957 term is the erosion of the west that there may be social issues that are 357 00:19:35,040 --> 00:19:38,557 brewing in the west that may be hurting us from the inside. How much do you 358 00:19:38,640 --> 00:19:41,117 observe or spend time on these issues? 359 00:19:41,200 --> 00:19:45,597 And the metric that often is cited now is declining birth rates in the west. 360 00:19:45,680 --> 00:19:48,717 And that our population, and we're going to talk with Elon in a few minutes about 361 00:19:48,800 --> 00:19:53,517 this. Um, oh, sorry. >> Oh, we just ran. My >> bad. My surprise. Oops. 362 00:19:53,600 --> 00:19:55,837 >> Sorry. Sorry. >> There's your surprise guest. 363 00:19:55,920 --> 00:19:57,917 >> Sorry. Sorry. Sorry. Sorry. 364 00:19:58,000 --> 00:20:02,637 >> Um, slip. Uh, but um 365 00:20:02,720 --> 00:20:04,237 >> Elon is a good friend and he's 366 00:20:04,320 --> 00:20:09,850 addressing this issue of population directly himself. [laughter] 367 00:20:11,680 --> 00:20:13,277 problem solve for it. Good for you. 368 00:20:13,360 --> 00:20:16,637 >> Is it is it is it a reflection of something going on? There's a rise of 369 00:20:16,720 --> 00:20:18,797 Mani getting elected in New York. Uh 370 00:20:18,880 --> 00:20:22,237 some of the historic values of the West 371 00:20:22,320 --> 00:20:25,357 seem to be un you know kind of under a a 372 00:20:25,440 --> 00:20:27,117 state of transformation right now. 373 00:20:27,200 --> 00:20:32,317 >> One metric of the success of a society is its ability to reproduce. And so I 374 00:20:32,400 --> 00:20:36,717 think this is a legitimate concern of the west. It's much worse in Asia. The 375 00:20:36,800 --> 00:20:39,117 um the Chinese number is about 1.0 0 for 376 00:20:39,200 --> 00:20:42,317 two parents. In Korea, it's now down to 377 00:20:42,400 --> 00:20:46,397 78 for two. So, it's really important to 378 00:20:46,480 --> 00:20:48,477 recognize that we as humans are 379 00:20:48,560 --> 00:20:51,197 collectively choosing to depopulate. And 380 00:20:51,280 --> 00:20:53,917 the numbers are staggering, right? And 381 00:20:54,000 --> 00:20:58,957 imagine a situation where instead of having growth, you have shrinkage. And 382 00:20:59,040 --> 00:21:03,357 furthermore, they're getting older. And so, as a business, all of a sudden, your 383 00:21:03,440 --> 00:21:06,477 revenue is declining. And there's nothing you can do because you can't 384 00:21:06,560 --> 00:21:08,877 innovate with fewer and fewer customers. 385 00:21:08,960 --> 00:21:11,997 So if you just put it in a business context, ignoring the moral issues which 386 00:21:12,080 --> 00:21:14,477 are all very real, it's just bad, right? 387 00:21:14,560 --> 00:21:17,677 So we have to solve that problem. I happen to be in favor broadly of 388 00:21:17,760 --> 00:21:21,277 immigration because I think immigration helps us solve that problem. But as a 389 00:21:21,360 --> 00:21:23,277 global mechanism, we have to address 390 00:21:23,360 --> 00:21:26,397 that. Um, in any case, from my 391 00:21:26,480 --> 00:21:28,717 perspective, you're going to have these 392 00:21:28,800 --> 00:21:32,077 issues, but America is organized around 393 00:21:32,160 --> 00:21:34,957 the concept of American exceptionalism. 394 00:21:35,040 --> 00:21:37,437 And as long as we understand that the 395 00:21:37,520 --> 00:21:40,477 way we make progress is we invest in the 396 00:21:40,560 --> 00:21:45,117 right people, in the right businesses, we have a a strong capital market, we 397 00:21:45,200 --> 00:21:46,797 invest in the infrastructure that they 398 00:21:46,880 --> 00:21:49,677 need, um, we'll be fine. That is my 399 00:21:49,760 --> 00:21:54,157 actual opinion. Can can we go back to um AI for a second? 400 00:21:54,240 --> 00:21:59,357 >> So um Eric, I think you can help us get to a let's call it a a bipartisan 401 00:21:59,440 --> 00:22:02,557 understanding of these issues because I think you you you think really clearly 402 00:22:02,640 --> 00:22:06,317 about this. um you know in in the wake 403 00:22:06,400 --> 00:22:09,917 of chat GBT launching at the end of 2022 404 00:22:10,000 --> 00:22:11,917 I think the discourse was really 405 00:22:12,000 --> 00:22:15,037 dominated in 2023 and 24 by this idea of 406 00:22:15,120 --> 00:22:17,997 AGI and that AGI was imminent and I 407 00:22:18,080 --> 00:22:22,077 think it created almost like a panicky atmosphere in Washington among policy 408 00:22:22,160 --> 00:22:24,637 makers and you saw things like we got to 409 00:22:24,720 --> 00:22:29,357 restrict open source because you know then China will get it and um and this 410 00:22:29,440 --> 00:22:32,157 is before DeepS launched and then we saw that actually they're ahead of us in 411 00:22:32,240 --> 00:22:34,797 open source But it feels like there's 412 00:22:34,880 --> 00:22:39,437 been um a pullback a little bit from the AGI narrative which I think I think is 413 00:22:39,520 --> 00:22:43,837 actually a good thing. I think it's more conducive to calm rational policym. 414 00:22:43,920 --> 00:22:45,677 What's your perception of AGI right now? 415 00:22:45,760 --> 00:22:47,917 Where where are we on that whole train? 416 00:22:48,000 --> 00:22:51,197 >> So so um so first place the speech that 417 00:22:51,280 --> 00:22:56,637 the president delivered about a month ago about AI strategy which I think you 418 00:22:56,720 --> 00:23:00,877 probably wouldn't say it but you kind of wrote it for him was exactly right. 419 00:23:00,960 --> 00:23:04,317 Right. So, thank you. >> David collaborated with an amazing 420 00:23:04,400 --> 00:23:07,677 leader who we all respect and admire so 421 00:23:07,760 --> 00:23:10,157 much. Eric, 422 00:23:10,240 --> 00:23:15,197 >> yes. Uh, so nevertheless, >> saying I wrote it was was way too 423 00:23:15,280 --> 00:23:18,237 strong. I mean, actually, but anyway, >> if you didn't if you didn't write it, 424 00:23:18,320 --> 00:23:20,077 then it must have been your twin. But in 425 00:23:20,160 --> 00:23:23,517 any case, um the you got you got the 426 00:23:23,600 --> 00:23:28,237 emphasis right, which was that investment in research, investment in 427 00:23:28,320 --> 00:23:32,317 the kind of stuff that we do is really really important. 428 00:23:32,400 --> 00:23:36,477 I don't agree with you on this on this AGI thing because there's this group 429 00:23:36,560 --> 00:23:39,277 which I call the San Francisco um 430 00:23:39,360 --> 00:23:43,517 narrative because they all live in San Francisco and their narrative goes 431 00:23:43,600 --> 00:23:48,237 something like this. Um today we're doing agents. Uh the agentic revolution 432 00:23:48,320 --> 00:23:52,717 will change businesses which I agree with. Um that what happens is the 433 00:23:52,800 --> 00:23:56,877 systems will become recursively self-intelligent. 434 00:23:56,960 --> 00:24:01,597 With recursive self-improvement as it's called if you have a scale-free problem 435 00:24:01,680 --> 00:24:05,837 and a scale-free problem for example is programming or math where you can just 436 00:24:05,920 --> 00:24:08,557 keep doing it. You get these enormous 437 00:24:08,640 --> 00:24:13,117 fast gains if you buy enough hardware or do enough software so forth and so on. 438 00:24:13,200 --> 00:24:15,517 that is still underway. 439 00:24:15,600 --> 00:24:18,477 The collective of that says that in the 440 00:24:18,560 --> 00:24:21,517 next three-ish years they believe that 441 00:24:21,600 --> 00:24:24,237 we will get forms of super intelligence 442 00:24:24,320 --> 00:24:27,277 and the way they define it is basically 443 00:24:27,360 --> 00:24:30,157 a savant a chemist soant a physics soant 444 00:24:30,240 --> 00:24:34,397 a mathematician soant I don't agree with the three years but I do agree that 445 00:24:34,480 --> 00:24:39,037 it'll be maybe six or seven years >> but if it's a savant in you know a 446 00:24:39,120 --> 00:24:42,077 particular area is that general intelligence 447 00:24:42,160 --> 00:24:47,197 >> it's not general intelligence Yet, general intelligence is when it can set 448 00:24:47,280 --> 00:24:49,517 its own objective function. >> Right? 449 00:24:49,600 --> 00:24:51,117 >> And >> there's no evidence of that. 450 00:24:51,200 --> 00:24:54,237 >> There's no evidence right now of the ability to set your own objective 451 00:24:54,320 --> 00:24:56,797 function. Um the the the thinking and 452 00:24:56,880 --> 00:24:58,717 I'm writing a paper on this so I've been 453 00:24:58,800 --> 00:25:01,837 studying it is that the the the 454 00:25:01,920 --> 00:25:06,797 technical problem is non-stationerity of mathematical proofs. And that what 455 00:25:06,880 --> 00:25:10,237 you're doing is you're trying to solve against objective function, but the 456 00:25:10,320 --> 00:25:14,317 objective function keeps changing, which is how humans operate. your goal changes 457 00:25:14,400 --> 00:25:15,997 every day. Whereas computers have 458 00:25:16,080 --> 00:25:18,557 trouble with that. As a math problem, we 459 00:25:18,640 --> 00:25:21,517 don't have an algorithm yet for LLMs 460 00:25:21,600 --> 00:25:23,437 that can do that. People are working on 461 00:25:23,520 --> 00:25:27,517 it. Um and the the test will be can you 462 00:25:27,600 --> 00:25:30,157 basically um using the information 463 00:25:30,240 --> 00:25:33,357 available in 1902, can you derive the 464 00:25:33,440 --> 00:25:35,037 same thing that Einstein did with 465 00:25:35,120 --> 00:25:37,277 special relativity followed by general 466 00:25:37,360 --> 00:25:40,797 relativity? We cannot do that today. Um 467 00:25:40,880 --> 00:25:43,677 and most people believe that the way 468 00:25:43,760 --> 00:25:46,237 this will be solved is through analogy. 469 00:25:46,320 --> 00:25:49,197 So the theory of great geniuses is that 470 00:25:49,280 --> 00:25:51,917 they understand one area extremely well 471 00:25:52,000 --> 00:25:54,877 and they're so brilliant the lady or man 472 00:25:54,960 --> 00:25:56,877 can then take their ideas and apply it 473 00:25:56,960 --> 00:26:00,157 to a completely different domain. If we 474 00:26:00,240 --> 00:26:02,717 can solve that problem then I think it's 475 00:26:02,800 --> 00:26:06,237 over. Then we get to AGI and then it's a 476 00:26:06,320 --> 00:26:07,917 whole different world. 477 00:26:08,000 --> 00:26:10,237 I think one of the reasons why it's hard 478 00:26:10,320 --> 00:26:14,317 to replace a human and you know JK and I 479 00:26:14,400 --> 00:26:18,637 debate this is that humans are end to end. You know we can do the whole job. 480 00:26:18,720 --> 00:26:21,357 You have sort of a complete understanding. You can pivot very 481 00:26:21,440 --> 00:26:25,677 easily. AI at least as we know it today is not end to end. It has to be 482 00:26:25,760 --> 00:26:29,517 prompted. You get an answer. That answer has to be validated. Then you have to 483 00:26:29,600 --> 00:26:32,477 ask a new question because it never gives you exactly what you want. You 484 00:26:32,560 --> 00:26:35,517 have to apply more context. You have to go through an iterative loop. Finally, 485 00:26:35,600 --> 00:26:39,997 you get to an answer that has business value. The way biology puts it is that 486 00:26:40,080 --> 00:26:41,917 AI is not endto end. It's middle to 487 00:26:42,000 --> 00:26:44,077 middle. Humans are endtoend. And so, as 488 00:26:44,160 --> 00:26:47,037 a result of that, instead of AI 489 00:26:47,120 --> 00:26:49,677 replacing all of us, AI will be very 490 00:26:49,760 --> 00:26:54,157 synergistic with humans because we can define the objective function. We do the 491 00:26:54,240 --> 00:26:58,237 prompting and we work with it to iterate and it does a lot of the work in the 492 00:26:58,320 --> 00:27:01,037 middle. Um, that seems to me like a very 493 00:27:01,120 --> 00:27:04,317 optimistic, less doomeristic take on it. 494 00:27:04,400 --> 00:27:08,717 What you just said is exactly what's going to happen for the next few years 495 00:27:08,800 --> 00:27:11,197 that each of us will have assistance 496 00:27:11,280 --> 00:27:13,677 which on our command and our prompting 497 00:27:13,760 --> 00:27:18,397 will be incredibly helpful to whatever problem we have you know personal uh you 498 00:27:18,480 --> 00:27:21,357 have people who are using these things for relationship advice for you know 499 00:27:21,440 --> 00:27:25,037 talking to their kids I mean it's all crazy stuff um but the fact of the 500 00:27:25,120 --> 00:27:29,437 matter is that's it the to me the real question is when does it cross over to 501 00:27:29,520 --> 00:27:33,437 having its own valition its own ability to seek information and solve new 502 00:27:33,520 --> 00:27:35,197 problems s that's a different animal. 503 00:27:35,280 --> 00:27:38,557 >> But have we seen any evidence of recursive self-improvement yet? 504 00:27:38,640 --> 00:27:41,597 >> Um, not yet. I'm I'm I've funded a 505 00:27:41,680 --> 00:27:45,837 number of startups which claim to be close to it, but of course these are 506 00:27:45,920 --> 00:27:50,557 startups and you never know, which tells me it's 5 10 years cutting numbers. 507 00:27:50,640 --> 00:27:52,797 >> What do you think Google's doing on this front? 508 00:27:52,880 --> 00:27:55,837 >> Um, well, I'm not at Google anymore. Um, 509 00:27:55,920 --> 00:27:58,877 every issue of Gemini is top of the 510 00:27:58,960 --> 00:28:02,237 leaderboard. So 2.5 just overcame 511 00:28:02,320 --> 00:28:06,797 everybody. and I'm sure there's another one coming. Um, Demis is working really 512 00:28:06,880 --> 00:28:08,797 hard on this question about um, 513 00:28:08,880 --> 00:28:11,757 scientific discovery. So, that's a p 514 00:28:11,840 --> 00:28:15,677 that is a path to getting to AGI. 515 00:28:15,760 --> 00:28:19,597 >> Eric, um, we appreciate the work you're doing. Uh, we appreciate you being here 516 00:28:19,680 --> 00:28:23,357 with us. We appreciate what you've done, the impact you've had on Silicon Valley, 517 00:28:23,440 --> 00:28:26,957 uh, as society. Yeah. No, but it's it's really been 518 00:28:27,040 --> 00:28:31,277 >> I am so happy to be part of this. You created this incredible community and 519 00:28:31,360 --> 00:28:36,037 there's all of these smart people that spend all their time listening to you. 520 00:28:36,120 --> 00:28:40,680 [laughter] >> Very concerning. 521 00:28:40,800 --> 00:28:44,600 >> Wow. Eric. 522 00:28:45,940 --> 00:28:47,960 [music] 523 00:28:49,840 --> 00:28:51,757 Thanks, Eric. Appreciate you. Cheers. 524 00:28:51,840 --> 00:28:56,200 [music] All right.