1 00:00:00,000 --> 00:00:03,837 We're living through a historic moment right now. The next thing that's really 2 00:00:03,920 --> 00:00:07,637 interesting and terrifying also is recursive self-improvement, but we don't 3 00:00:07,720 --> 00:00:10,637 have it yet. What we do have 4 00:00:10,720 --> 00:00:13,277 I keep asking my friends, when does the 5 00:00:13,360 --> 00:00:15,637 asymptote arrive and when does the curve 6 00:00:15,720 --> 00:00:17,877 slow down? It is actually true that 7 00:00:17,960 --> 00:00:20,637 there is a limit to our craziness. 8 00:00:20,720 --> 00:00:23,517 We have not found it yet and that's the 9 00:00:23,600 --> 00:00:25,917 great thing frankly about America. 10 00:00:26,000 --> 00:00:28,357 The American competitor, not enemy, but 11 00:00:28,440 --> 00:00:30,557 competitor is China. They have lots of 12 00:00:30,640 --> 00:00:33,277 money. They're very, very smart. Their 13 00:00:33,360 --> 00:00:35,637 work ethic is equal or stronger than 14 00:00:35,720 --> 00:00:38,637 ours and they dominate key industries. 15 00:00:38,720 --> 00:00:42,957 But at the moment it sure looks to me like the robotic hardware 16 00:00:43,040 --> 00:00:45,517 of China is the winner. I don't want to 17 00:00:45,600 --> 00:00:48,837 lose the robotic revolution in my view 18 00:00:48,920 --> 00:00:51,077 the way we lost the electric vehicle 19 00:00:51,160 --> 00:00:57,160 revolution at least on the low end. It's possible, but it requires 20 00:00:57,680 --> 00:01:01,680 Now that's a moon shot, ladies and gentlemen. 21 00:01:02,640 --> 00:01:05,317 Eric, do you remember the first time we met? 22 00:01:05,400 --> 00:01:07,397 Yes, Larry Page introduced me to you 23 00:01:07,480 --> 00:01:08,677 because he was on your board. 24 00:01:08,760 --> 00:01:12,477 >> Yes. Yeah, so I got a call from Eric out 25 00:01:12,560 --> 00:01:15,197 of the blue, which was a great honor. 26 00:01:15,280 --> 00:01:19,117 And and said Larry says I should meet you. When are you 27 00:01:19,200 --> 00:01:22,437 going to be down here in or up here in San Francisco? I was in LA 28 00:01:22,520 --> 00:01:24,417 and I said how about tomorrow? 29 00:01:24,500 --> 00:01:27,037 >> [laughter] >> Typical Peter. 30 00:01:27,120 --> 00:01:29,677 And I remember something which was so 31 00:01:29,760 --> 00:01:32,477 I love the story. We sat down to lunch 32 00:01:32,560 --> 00:01:34,317 at Charlie's Cafe. 33 00:01:34,400 --> 00:01:39,717 Um and of course I'm running a nonprofit and my mission is always raising capital 34 00:01:39,800 --> 00:01:43,317 for nonprofit. And so we sit down to lunch 35 00:01:43,400 --> 00:01:47,477 and before we get started you said Peter, so what is your highest level of 36 00:01:47,560 --> 00:01:51,637 of giving a membership? And I said well Eric, it's our vision circle 37 00:01:51,720 --> 00:01:55,857 for $2.5 million and you go, okay, I'm in. Now let's have a conversation. 38 00:01:55,940 --> 00:02:00,437 >> [laughter] >> I'll buy them all. It was it was crazy. 39 00:02:00,520 --> 00:02:03,277 >> [laughter] >> Yeah. Oh, there you go. 40 00:02:03,360 --> 00:02:08,077 But you know, somehow you have the reason when we 41 00:02:08,160 --> 00:02:11,637 spoke, the reason I wanted to come here is this has become the epicenter of the 42 00:02:11,720 --> 00:02:16,117 abundance movement. And the abundance movement is correct. That's the 43 00:02:16,200 --> 00:02:18,777 important thing. Thank you. >> Yeah. 44 00:02:18,860 --> 00:02:20,880 >> [applause] 45 00:02:21,040 --> 00:02:23,597 >> Um I just want to thank you for all the 46 00:02:23,680 --> 00:02:26,677 support you've given myself, X Prize all 47 00:02:26,760 --> 00:02:29,557 these years. So grateful for that. So 48 00:02:29,640 --> 00:02:33,197 I'll start with a a question that I'd 49 00:02:33,280 --> 00:02:37,237 love to hear you expand on, which is we're living through a historic moment 50 00:02:37,320 --> 00:02:39,997 right now. Right? Um 51 00:02:40,080 --> 00:02:41,837 could you define the moment that we're 52 00:02:41,920 --> 00:02:44,557 in and give us sort of a a state of the 53 00:02:44,640 --> 00:02:48,317 union of what's going on in AI? 54 00:02:48,400 --> 00:02:52,437 We're 10 or 15% into the impacts of 55 00:02:52,520 --> 00:02:55,757 this. And you can see it. 56 00:02:55,840 --> 00:03:00,677 You can feel it. And some of it will happen, some of it 57 00:03:00,760 --> 00:03:04,557 will take longer, right? So for example, hardware takes longer than software. So 58 00:03:04,640 --> 00:03:09,317 that robots take longer than digital systems on traditional hardware, things 59 00:03:09,400 --> 00:03:13,917 like that. Um we've not um 60 00:03:14,000 --> 00:03:18,037 the next thing that's really interesting and terrifying also is recursive 61 00:03:18,120 --> 00:03:21,757 self-improvement. >> Mhm. It's not happening yet. 62 00:03:21,840 --> 00:03:26,277 And so it's easy to convince yourself that you're going to have human agents 63 00:03:26,360 --> 00:03:29,517 sorry, computer agents that are human-like completely within a year or 64 00:03:29,600 --> 00:03:32,037 two. We don't have the science for that 65 00:03:32,120 --> 00:03:34,277 yet. People are working on it. I can 66 00:03:34,360 --> 00:03:37,237 describe how I think it'll play out. But 67 00:03:37,320 --> 00:03:39,837 we don't have it yet. What we do have is 68 00:03:39,920 --> 00:03:42,677 reasoning systems that are perfect 69 00:03:42,760 --> 00:03:46,917 partners for human beings for good and bad, right? And that has a lot of 70 00:03:47,000 --> 00:03:51,197 implications. So if we stop today, which we're not and it's not stoppable or 71 00:03:51,280 --> 00:03:55,037 controllable by any government or any single individual or corporation, 72 00:03:55,120 --> 00:03:59,477 we would still have advanced humanity because of these reasoning agents. How 73 00:03:59,560 --> 00:04:03,437 fast do you imagine this is going to accelerate? 74 00:04:03,520 --> 00:04:07,197 Um the there's a thing which I call the San 75 00:04:07,280 --> 00:04:11,077 Francisco consensus and the reason I call it is cuz everyone in San Francisco 76 00:04:11,160 --> 00:04:12,797 believes this. 77 00:04:12,880 --> 00:04:15,637 Uh everyone I know anyway, which is that 78 00:04:15,720 --> 00:04:20,117 it it it's easy to understand. Uh this is the year of agents, which we can 79 00:04:20,200 --> 00:04:23,837 discuss why agents will take over everything this year. 80 00:04:23,920 --> 00:04:27,637 During this year the scaling of the use of agents and reasoning will sort of 81 00:04:27,720 --> 00:04:30,637 grow at this enormous rate. Everybody's out of 82 00:04:30,720 --> 00:04:32,557 hardware, everyone's out of electricity. 83 00:04:32,640 --> 00:04:35,917 It's a real boom, right? It's like the biggest boom I've seen and I've been 84 00:04:36,000 --> 00:04:39,837 through three or four of these in my career. 85 00:04:39,920 --> 00:04:43,677 Um the in this thinking 86 00:04:43,760 --> 00:04:47,397 once you have recursive self-improvement where the system can begin to improve 87 00:04:47,480 --> 00:04:49,837 itself, you have intelligence learning 88 00:04:49,920 --> 00:04:54,517 on its own and it will in this argument, it will learn faster than we can cuz 89 00:04:54,600 --> 00:04:58,877 we're biologically limited. And the way this is expressed in San 90 00:04:58,960 --> 00:05:00,557 Francisco and I'll give a simple 91 00:05:00,640 --> 00:05:02,237 example. You have a tech company with a 92 00:05:02,320 --> 00:05:07,277 thousand fantastic AI researchers. Um 93 00:05:07,360 --> 00:05:10,717 so one day they turn on AI research that 94 00:05:10,800 --> 00:05:13,437 is an AI research agent. Well, how many 95 00:05:13,520 --> 00:05:18,877 AI research agents do you have? Well, as many as you're limited by electricity, 96 00:05:18,960 --> 00:05:21,237 right? You don't have to feed them, they don't 97 00:05:21,320 --> 00:05:23,637 need housing, there's no more housing in San Francisco, you know, all that kind 98 00:05:23,720 --> 00:05:26,437 of stuff. Right? You don't have those problems. You don't have an HR 99 00:05:26,520 --> 00:05:29,957 department for them, if you will. Uh and you don't have to pay them, you just 100 00:05:30,040 --> 00:05:33,437 have to feed them electricity. So how many could you have? Well, maybe a maybe 101 00:05:33,520 --> 00:05:35,717 a million of these agents. 102 00:05:35,800 --> 00:05:38,997 Now in AI the way you determine you've 103 00:05:39,080 --> 00:05:42,557 made progress is you have clear metrics 104 00:05:42,640 --> 00:05:44,197 that the reasoning or testing or 105 00:05:44,280 --> 00:05:46,317 whatever the evaluation framework is 106 00:05:46,400 --> 00:05:50,357 better, right? So that's what happens. 107 00:05:50,440 --> 00:05:53,877 So in that scenario the slope goes like 108 00:05:53,960 --> 00:05:58,237 this because you're already at this slope, then you add more people, then 109 00:05:58,320 --> 00:06:02,597 you get the agents and you go like this and this is essentially 110 00:06:02,680 --> 00:06:04,837 a super intelligence moment. 111 00:06:04,920 --> 00:06:06,597 The belief in San Francisco is this 112 00:06:06,680 --> 00:06:09,877 occurs within two to three years. 113 00:06:09,960 --> 00:06:13,117 The evidence in favor 114 00:06:13,200 --> 00:06:17,357 goes something like this. Claude code came out a couple months 115 00:06:17,440 --> 00:06:20,277 ago, the the latest one, Opus whatever it is. What was it? 116 00:06:20,360 --> 00:06:23,037 >> 1.6, yes, thank you. And 117 00:06:23,120 --> 00:06:28,797 everyone I know in the Bay Area that's doing software says it was 80/20, now 118 00:06:28,880 --> 00:06:30,637 it's 20/80. >> Mhm. 119 00:06:30,720 --> 00:06:34,157 The best analysis I can come up with is 120 00:06:34,240 --> 00:06:36,837 it's not the Claude code part, it's that 121 00:06:36,920 --> 00:06:41,117 the that the underlying LLM can produce 122 00:06:41,200 --> 00:06:43,637 more reasoning over time, better quality 123 00:06:43,720 --> 00:06:46,317 tokens over time. It's a deeper thinker, 124 00:06:46,400 --> 00:06:50,757 Mhm. right? And all the labs now are competing for that. This is not just the 125 00:06:50,840 --> 00:06:53,597 size of the context window, it's actually the reasoning skill and the 126 00:06:53,680 --> 00:06:57,997 length of which it can think. It can just think longer and produce more 127 00:06:58,080 --> 00:07:01,957 stuff, right? I watched this stuff when I was I 128 00:07:02,040 --> 00:07:05,717 moved to the the Bay Area when I was 21 129 00:07:05,800 --> 00:07:09,037 and I was a programmer in high school way back when. 130 00:07:09,120 --> 00:07:10,837 And I was pretty good programmer. 131 00:07:10,920 --> 00:07:13,037 And I watch what it does and I go, my 132 00:07:13,120 --> 00:07:16,197 god, I'm over, you know? There's not a 133 00:07:16,280 --> 00:07:19,157 thing that I could do that it cannot do. 134 00:07:19,240 --> 00:07:22,637 So when they wrote a C compiler 135 00:07:22,720 --> 00:07:26,077 in rust I got it's over, you know, like 136 00:07:26,160 --> 00:07:29,837 declare suc So I think [laughter] part of this part of this is because the 137 00:07:29,920 --> 00:07:33,917 people who are building it are also seeing the diminution of their own 138 00:07:34,000 --> 00:07:38,197 skill. They're being forced to go from programmers, which is what I'm very 139 00:07:38,280 --> 00:07:42,957 proud to have been, to being the the director of a programming system, 140 00:07:43,040 --> 00:07:46,997 Mhm. >> right? And the most likely scenario, by 141 00:07:47,080 --> 00:07:49,197 the way, there's a lot of implications for this. 142 00:07:49,280 --> 00:07:53,197 One is that it's always been true, speaking as your local arrogant 143 00:07:53,280 --> 00:07:57,917 programmer, that the very top programmers were worth 10 times more 144 00:07:58,000 --> 00:07:59,957 than the ones right below. There's something 145 00:08:00,040 --> 00:08:03,597 >> special about the mathematical reasoning skills of programmers. 146 00:08:03,680 --> 00:08:06,397 Those people will become more valuable, 147 00:08:06,480 --> 00:08:10,717 not less valuable, because these systems need to be controlled by humans at the 148 00:08:10,800 --> 00:08:14,517 moment. Those people will be capable of grasping the parallelization and 149 00:08:14,600 --> 00:08:17,557 activities of this. It also means that 150 00:08:17,640 --> 00:08:21,797 you're going to have, in my view, a relatively small number of very large 151 00:08:21,880 --> 00:08:24,597 companies. >> Yeah. Yeah. And this is a big deal. 152 00:08:24,680 --> 00:08:28,357 >> Yeah. Yeah. And a very large number of very small companies cuz you don't need 153 00:08:28,440 --> 00:08:30,637 as many people. And you're watching that 154 00:08:30,720 --> 00:08:35,717 play out this month. Mhm. I mean this all happened in the last three I got I 155 00:08:35,800 --> 00:08:38,317 one I was in one startup I'm involved with. 156 00:08:38,400 --> 00:08:41,637 I was talking to the programmer who was, you know, perfectly perfectly brilliant 157 00:08:41,720 --> 00:08:44,157 young man. And I said, well, what's the 158 00:08:44,240 --> 00:08:46,397 truth? He said, well, here's what I do. 159 00:08:46,480 --> 00:08:50,757 Um he's working on UIs of various kinds and I said, I write the spec of what I 160 00:08:50,840 --> 00:08:53,037 want and then I write a test function, 161 00:08:53,120 --> 00:08:55,877 an evaluation function Mhm. and then I 162 00:08:55,960 --> 00:09:00,397 turn it on. I said, what time? And he goes, 7:00 163 00:09:00,480 --> 00:09:03,397 in the evening. And I go, okay, what do you then do? 164 00:09:03,480 --> 00:09:06,957 Well, he has dinner with his wife and he goes to sleep. And I said, do you wake 165 00:09:07,040 --> 00:09:09,467 up? He said, no, I sleep very well. 166 00:09:09,550 --> 00:09:11,957 >> [laughter] >> When does it finish? Oh, 4:00 in the 167 00:09:12,040 --> 00:09:16,557 morning. Mhm. And then he gets up, has breakfast, 168 00:09:16,640 --> 00:09:20,117 you know, does whatever he does and then he sees what's been invented. I mean 169 00:09:20,200 --> 00:09:23,237 it's mind-boggling. And and the stupid 170 00:09:23,320 --> 00:09:27,597 example I used with this young man, this is what the power of these systems 171 00:09:27,680 --> 00:09:32,677 are. If you can define the evaluation function and you can let it run and if 172 00:09:32,760 --> 00:09:36,717 you have enough hardware, you're inventing worlds. I mean this stuff 173 00:09:36,800 --> 00:09:39,797 would have taken me 6 months and 10 programmers at Google to do the same 174 00:09:39,880 --> 00:09:42,677 thing and this poor guy's sleeping. It's so funny you say that cuz I was 175 00:09:42,760 --> 00:09:46,877 literally backstage, they said Eric Schmidt's coming now and I have my lid 176 00:09:46,960 --> 00:09:51,037 open on my Mac and I'm trying to get the jobs onto the 177 00:09:51,120 --> 00:09:54,877 cloud so I can close the lid cuz if you close the lid it'll break the jobs. So 178 00:09:54,960 --> 00:09:58,477 I've got these six concurrent Claude 4.6 >> open. 179 00:09:58,560 --> 00:10:01,237 >> can you know, it's like important what you're doing. 180 00:10:01,320 --> 00:10:05,037 Tony interrupt for me. You're important, too, you know. 181 00:10:05,120 --> 00:10:09,917 No, it it's it's crazy cuz when we got together in Davos just what 2 months ago, 182 00:10:10,000 --> 00:10:13,797 uh it was in this kind of autocomplete mode. You'd write the code and then it 183 00:10:13,880 --> 00:10:16,357 would help you get it done. You're about 10 times more efficient, but you're 184 00:10:16,440 --> 00:10:21,077 still babysitting it. Now, literally it's working right now. When I 185 00:10:21,160 --> 00:10:25,397 get off stage, it will have solved six problems that I launched. And And and I 186 00:10:25,480 --> 00:10:28,877 appreciate the excitement in the industry, but I can tell you when I was 187 00:10:28,960 --> 00:10:32,557 I used to work on BSD. I basically worked at Berkeley on Unix at Bell Labs 188 00:10:32,640 --> 00:10:34,157 and at BSD Unix. 189 00:10:34,240 --> 00:10:37,197 And we programmers invent what they 190 00:10:37,280 --> 00:10:42,357 need. And so we invented the first email system and the first messaging system. 191 00:10:42,440 --> 00:10:45,077 And nobody thought about it. It was like, well, we just need this thing. 192 00:10:45,160 --> 00:10:49,517 Yeah. So one key thing to understand about digital intelligence is the first 193 00:10:49,600 --> 00:10:52,757 inventors are the people who are solving problems of themselves who are 194 00:10:52,840 --> 00:10:56,517 programmers. Mhm. Right? You shouldn't be surprised by this. You should have 195 00:10:56,600 --> 00:10:59,917 expected this. The other thing that's interesting about 196 00:11:00,000 --> 00:11:03,677 programming is it's um it's both scale free, which means there's no no 197 00:11:03,760 --> 00:11:07,277 particular limitations except electricity. You don't need a lot of 198 00:11:07,360 --> 00:11:11,437 data and you already have GitHub and the equivalents. And it's also a fairly 199 00:11:11,520 --> 00:11:15,637 limited language set. So the number of language components, if you will, 200 00:11:15,720 --> 00:11:17,437 compared to human language is less. 201 00:11:17,520 --> 00:11:21,357 Smaller language, clear objective function, all you need is electricity. 202 00:11:21,440 --> 00:11:25,877 Yep. Now, how far can this go? It'll get to the point where you don't have the 203 00:11:25,960 --> 00:11:27,677 ability to do completely new things. 204 00:11:27,760 --> 00:11:31,637 Yeah. Isn't it Isn't it really quaint and crazy to think that we can sit here and 205 00:11:31,720 --> 00:11:34,277 say, yeah, I wrote a ton of code when I was younger. 206 00:11:34,360 --> 00:11:37,877 No one will ever do that again after the end of this year. It'll be like riding a 207 00:11:37,960 --> 00:11:41,437 horse, you know, be like quaint skills that we all used to have. No, but I do 208 00:11:41,520 --> 00:11:44,077 have a proposal for universities. Those of you who are associated with 209 00:11:44,160 --> 00:11:46,517 universities, you should stop everything else you're doing in the university 210 00:11:46,600 --> 00:11:51,237 right now and design a course for freshman men and women 211 00:11:51,320 --> 00:11:54,997 uh uh who starting in September, 212 00:11:55,080 --> 00:11:57,997 which is a prompt engineering class. 213 00:11:58,080 --> 00:12:01,437 Why Why Why university? Why not high school? 214 00:12:01,520 --> 00:12:03,697 God, you're so aggressive, Peter. 215 00:12:03,780 --> 00:12:06,477 >> [laughter] >> Let's start with universities. You can 216 00:12:06,560 --> 00:12:10,037 improve my idea. I I thought 18-year-olds would be would be young 217 00:12:10,120 --> 00:12:12,357 enough. Maybe you think it's younger. 218 00:12:12,440 --> 00:12:16,997 Here's the most important thing. Spend Spend Spend a quarter or a semester. The 219 00:12:17,080 --> 00:12:21,277 first thing they learn in university is how to use these tools. 220 00:12:21,360 --> 00:12:24,197 Universities are completely opposed to my idea, as usual. 221 00:12:24,280 --> 00:12:28,277 >> Yeah. Because it violates every one of their tenets. But if you think about the 222 00:12:28,360 --> 00:12:30,517 student, and I mean every student, 223 00:12:30,600 --> 00:12:33,037 liberal arts, you know, math, whatever, 224 00:12:33,120 --> 00:12:35,757 they're going this platform will be the 225 00:12:35,840 --> 00:12:38,517 expression platform for their art, their 226 00:12:38,600 --> 00:12:41,117 music, their writing, and so forth. Why 227 00:12:41,200 --> 00:12:43,037 wouldn't you teach them immediately? 228 00:12:43,120 --> 00:12:46,037 Peter, improve my proposal. No, I I just 229 00:12:46,120 --> 00:12:49,117 I just feel like we're that AI is going 230 00:12:49,200 --> 00:12:52,997 to impact every student in high school 231 00:12:53,080 --> 00:12:56,757 today. Yeah. And that they're living an 232 00:12:56,840 --> 00:12:59,917 unnatural life by not engaging with it. 233 00:13:00,000 --> 00:13:04,517 And when they hit universities, for those that still exist, um 234 00:13:04,600 --> 00:13:08,037 Well, plus your kids are that age, and so they're literally right now doing 235 00:13:08,120 --> 00:13:11,237 exactly what you're describing. So >> Peo- People here who have teenagers, you 236 00:13:11,320 --> 00:13:13,797 know what I'm talking about cuz they're all in it already. So So I think you're 237 00:13:13,880 --> 00:13:15,517 that's an improvement to my argument. 238 00:13:15,600 --> 00:13:19,557 The- There's a problem of age restriction. Uh you really have to think 239 00:13:19,640 --> 00:13:23,237 about vulnerable teenagers with this technology. Um >> [clears throat] 240 00:13:23,320 --> 00:13:25,077 >> I did some some analysis of where the 241 00:13:25,160 --> 00:13:27,317 real problems are with this stuff. 242 00:13:27,400 --> 00:13:30,797 Um a simple summary is that at some 243 00:13:30,880 --> 00:13:35,917 point there will be jobs impact from this stuff. We're seeing it in software 244 00:13:36,000 --> 00:13:39,077 and we're seeing it in certain customer service industries, not across the 245 00:13:39,160 --> 00:13:40,837 board. At some point that will happen. 246 00:13:40,920 --> 00:13:42,277 That's an issue. 247 00:13:42,360 --> 00:13:47,197 Another one is how do we as a country 248 00:13:47,280 --> 00:13:49,237 maintain our moral values while we're 249 00:13:49,320 --> 00:13:51,717 also racing against China? Another one 250 00:13:51,800 --> 00:13:55,037 is the impact on um 251 00:13:55,120 --> 00:13:59,597 uh young people. It is not okay for 13-year-olds to be committing suicide 252 00:13:59,680 --> 00:14:02,557 because of an LLM. It's just not okay. It should 253 00:14:02,640 --> 00:14:05,997 It's It's It needs to be addressed. It needs to be addressed right now. For 254 00:14:06,080 --> 00:14:09,597 sure. Uh and there's all sorts of other issues. The other one I came up with was 255 00:14:09,680 --> 00:14:13,197 in agent orchestration. Agents can be combined. 256 00:14:13,280 --> 00:14:16,557 I've always been worried that when you put the agents together, 257 00:14:16,640 --> 00:14:20,677 especially if they're from non uh non-compatible vendors, you get 258 00:14:20,760 --> 00:14:26,037 unpredictable effects. Yeah. So So these are problems to be solved. So we herald 259 00:14:26,120 --> 00:14:28,877 the future and we solve the problems that it brought. So we're we're going to 260 00:14:28,960 --> 00:14:32,397 talk about China and government and jobs. 261 00:14:32,480 --> 00:14:36,397 But before we do that, um I want to say I'm I'm in this savor the moment's kind 262 00:14:36,480 --> 00:14:39,397 of mode right now because I feel like the world to a year from today will be 263 00:14:39,480 --> 00:14:41,957 nothing like the world today and all 264 00:14:42,040 --> 00:14:43,797 everything we're doing right now, 265 00:14:43,880 --> 00:14:46,637 I've enjoyed so much for so long and I 266 00:14:46,720 --> 00:14:51,117 just want to savor the moment, but reminisce for 1 minute about the fact 267 00:14:51,200 --> 00:14:55,037 that while you were running Google, the transformer got invented there. 268 00:14:55,120 --> 00:14:59,957 The TPU got invented there. Demis Hassabis solved protein folding, 269 00:15:00,040 --> 00:15:03,197 which is now universally used. You know, do Does the 270 00:15:03,280 --> 00:15:04,917 work in I think it's Does the work in an 271 00:15:05,000 --> 00:15:10,637 hour that used to take a PhD student 4 years 4 years. 272 00:15:10,720 --> 00:15:14,597 >> It's like 300 million times more efficient. Um all of that and all the 273 00:15:14,680 --> 00:15:17,437 diaspora from that, all the people working in the field in San Francisco, 274 00:15:17,520 --> 00:15:21,677 as you mentioned, they all were your people. Yeah. And so you were there at 275 00:15:21,760 --> 00:15:24,877 the creation of everything we're experiencing right now. Do you think 276 00:15:24,960 --> 00:15:28,117 anything like that profoundly strikes you about that moment? I mean, did you 277 00:15:28,200 --> 00:15:30,637 even realize at the time that it was still I think I think when you're making 278 00:15:30,720 --> 00:15:32,757 history, you typically don't know it. 279 00:15:32,840 --> 00:15:36,397 Um I give a lot of credit to Larry and Sergey. 280 00:15:36,480 --> 00:15:38,117 Because they were ahead of me. I'm I'm 281 00:15:38,200 --> 00:15:41,597 an operating CEO and they pushed pushed 282 00:15:41,680 --> 00:15:45,837 and pushed for excellence. And I'll give you an example. In the 283 00:15:45,920 --> 00:15:50,557 early years of Google, Larry would My favorite interaction was one day I 284 00:15:50,640 --> 00:15:54,237 said, we need to hire some people doing Java. 285 00:15:54,320 --> 00:15:58,837 And Larry and Sergey said, this is the stupidest idea we have ever heard. And 286 00:15:58,920 --> 00:16:01,517 And I could never tell with them whether they were being serious or not or they 287 00:16:01,600 --> 00:16:05,317 were just joking with me. But their argument was that real programmers were 288 00:16:05,400 --> 00:16:06,957 programming one level lower. Today 289 00:16:07,040 --> 00:16:09,117 Google has many thousands of of it. 290 00:16:09,200 --> 00:16:12,477 But they But they were so precise and so 291 00:16:12,560 --> 00:16:17,357 um driven to excellence in technology, Yeah. right? That I could not uh fool 292 00:16:17,440 --> 00:16:21,637 them. I couldn't market around them. I needed to have the technical expert. And 293 00:16:21,720 --> 00:16:25,077 they say, oh, that's boring. Don't do that. That's another one of your ideas. 294 00:16:25,160 --> 00:16:29,637 Right? We want a new idea. Wow. And I give them a lot of credit for it. 295 00:16:29,720 --> 00:16:31,077 >> But what about the TPU in particular? 296 00:16:31,160 --> 00:16:35,557 Like that to me cuz I I didn't even hear about it until much later. 297 00:16:35,640 --> 00:16:39,837 And it takes years to to build design and build your own internal chips. And 298 00:16:39,920 --> 00:16:44,157 now it's it's about to explode. I don't know how much is public or not, but it's 299 00:16:44,240 --> 00:16:48,597 it's just Well, so the TPU version one was essentially a matrix multiplier of a 300 00:16:48,680 --> 00:16:51,437 particular kind. When they went to version two, they changed the algorithm 301 00:16:51,520 --> 00:16:52,997 in a complicated way. And it's 302 00:16:53,080 --> 00:16:56,037 particularly good for inference. 303 00:16:56,120 --> 00:16:59,877 Whether it's brilliance or just luck, 304 00:16:59,960 --> 00:17:02,037 those decisions made 10 years ago set up 305 00:17:02,120 --> 00:17:04,477 the TPU as the perfect inference engine. 306 00:17:04,560 --> 00:17:09,717 And for everybody's benefit, inference is what the reasoning tech steps that 307 00:17:09,800 --> 00:17:12,797 I'm describing on. So Google is particularly well positioned. As you 308 00:17:12,880 --> 00:17:17,277 know, Nvidia purchased Groq um for the reason of getting that inner 309 00:17:17,360 --> 00:17:19,237 >> Yeah, trying to catch up to what you thought of 10 years ago. 310 00:17:19,320 --> 00:17:21,637 >> interesting about Nvidia, if you look at them, I was looking at the Rubin 311 00:17:21,720 --> 00:17:25,957 architecture. They managed to do Nvidia managed to do 312 00:17:26,040 --> 00:17:30,397 what Intel could never do. Intel could never get control of the complete server 313 00:17:30,480 --> 00:17:34,637 architecture. And they tried. Nvidia has managed to build real 314 00:17:34,720 --> 00:17:38,957 supercomputers that you can really buy with enough time and money and so forth. 315 00:17:39,040 --> 00:17:41,997 And it will really be delivered to your to your And And it just do the whole 316 00:17:42,080 --> 00:17:46,357 thing. These are major industrial achievements. And that's why both 317 00:17:46,440 --> 00:17:48,957 companies will do incredibly well. 318 00:17:49,040 --> 00:17:54,077 Right, for that reason. Eric, in the in the AI exponential growth right now, 319 00:17:54,160 --> 00:17:56,917 talk [clears throat] to me about where the constraints are. You were in 320 00:17:57,000 --> 00:18:00,317 Congress talking about energy, uh chips, 321 00:18:00,400 --> 00:18:04,197 um people, capital. Where are the constraints right now? Well, it's It's 322 00:18:04,280 --> 00:18:05,517 interesting. I started a data center 323 00:18:05,600 --> 00:18:11,720 company with my friends. Um there's a uh the 324 00:18:11,880 --> 00:18:13,637 In my testimony, I said there was an 325 00:18:13,720 --> 00:18:17,597 estimated 92 gigawatts shortage of power 326 00:18:17,680 --> 00:18:20,797 in America in the next between now and 2030. 327 00:18:20,880 --> 00:18:23,397 And uh by reference, a nuclear power 328 00:18:23,480 --> 00:18:27,917 plant is about 1.5 gigawatts. So it's about 60 nuclear plants and we're doing 329 00:18:28,000 --> 00:18:29,517 essentially zero or one, right? 330 00:18:29,600 --> 00:18:32,677 Depending on how you count. Um 331 00:18:32,760 --> 00:18:35,117 So I got interested in the question of 332 00:18:35,200 --> 00:18:40,397 what's the real resource constraint in in America? And it's electricity. Mhm. 333 00:18:40,480 --> 00:18:44,157 We have the universities. We have the smart people. We have the economics. And 334 00:18:44,240 --> 00:18:47,997 we have these amazing finance people who will give all of us billions and 335 00:18:48,080 --> 00:18:51,877 billions of dollars on a wing and a prayer. I mean, there's no country where 336 00:18:51,960 --> 00:18:55,557 the finance people are sufficiently crazy to do that. 337 00:18:55,640 --> 00:18:58,557 It's not true in China. It's not It's certainly not true in Europe. And these 338 00:18:58,640 --> 00:19:02,717 guys are incredibly jealous of the American financial system. So I always 339 00:19:02,800 --> 00:19:07,237 start Thank you to the finance people for funding our dreams whether they work 340 00:19:07,320 --> 00:19:09,317 or not, right? Thank [laughter] you. 341 00:19:09,400 --> 00:19:13,437 Uh >> [applause] >> So 342 00:19:13,520 --> 00:19:16,797 So the There There's usually a retort at this 343 00:19:16,880 --> 00:19:20,677 point where people say, well, the algorithms will ultimately require less 344 00:19:20,760 --> 00:19:23,917 energy. I am sure that that is true. 345 00:19:24,000 --> 00:19:28,477 There is this property that is the as the power as the power of the hardware 346 00:19:28,560 --> 00:19:30,517 goes up, as the algorithms 347 00:19:30,600 --> 00:19:35,557 become more efficient, you don't need less power, you need even more power and 348 00:19:35,640 --> 00:19:39,197 even more computers because we discover new uses. Yeah. Jevons paradox. 349 00:19:39,280 --> 00:19:40,917 >> It's called Jevons paradox. Yeah. 350 00:19:41,000 --> 00:19:44,757 And so I think that uh 351 00:19:44,840 --> 00:19:48,797 be- because humans have trouble with exponentials, everyone says, oh, well, 352 00:19:48,880 --> 00:19:51,877 in 6 to 9 months, it'll will the bubble and you know, so forth and so on. 353 00:19:51,960 --> 00:19:55,437 There's no sign of this. We've been I I I and a team have been 354 00:19:55,520 --> 00:20:00,117 working on this for years. The the ultimate law the ultimate 355 00:20:00,200 --> 00:20:04,197 essentially scaling laws are not done yet. 356 00:20:04,280 --> 00:20:06,757 I keep asking my friends, when does the 357 00:20:06,840 --> 00:20:08,637 asymptote arrive, and when does the 358 00:20:08,720 --> 00:20:11,637 curve slow down? We've not seen it yes 359 00:20:11,720 --> 00:20:14,397 yet. There will be one, right? 360 00:20:14,480 --> 00:20:18,677 It is actually true that there is a limit to our craziness. 361 00:20:18,760 --> 00:20:23,397 We have not found it yet. And we're running to the wall. And 362 00:20:23,480 --> 00:20:25,717 that's the great thing, frankly, about America. 363 00:20:25,800 --> 00:20:28,557 >> You think you think it's a limit to the capital, or a limit to if you just add 364 00:20:28,640 --> 00:20:32,837 more and more and more scale and parameters, something just doesn't work? 365 00:20:32,920 --> 00:20:35,437 Um Well, the first question is is there a 366 00:20:35,520 --> 00:20:38,797 limit to the capital available? 367 00:20:38,880 --> 00:20:42,997 Um a gigawatt of power corresponds to 368 00:20:43,080 --> 00:20:47,757 about 50 billion dollars of hardware, software, data centers. It's on the 369 00:20:47,840 --> 00:20:49,197 order of, depending on what numbers you 370 00:20:49,280 --> 00:20:53,357 use, okay? So, 100 gigawatts is do the 371 00:20:53,440 --> 00:20:56,957 math. Yeah. Um can we raise 5 trillion 372 00:20:57,040 --> 00:21:00,557 dollars over 5 years? Yeah. 373 00:21:00,640 --> 00:21:02,717 That's the strength of America. 374 00:21:02,800 --> 00:21:04,717 Could we double that? Mm, 375 00:21:04,800 --> 00:21:06,917 you know, we're we're we're getting the 376 00:21:07,000 --> 00:21:11,957 data center build out is 1% of the GDP of America growth right 377 00:21:12,040 --> 00:21:13,317 >> we're back to a power program. 378 00:21:13,400 --> 00:21:17,597 >> Right. Well, thank you. And the current estimate of electricity use 379 00:21:17,680 --> 00:21:21,877 um in America is that electricity the 10% of the electricity in the United 380 00:21:21,960 --> 00:21:23,837 States will be used in the data centers. 381 00:21:23,920 --> 00:21:26,837 And these data centers These are not the data centers I used to build at Google, 382 00:21:26,920 --> 00:21:30,677 which seemed tiny by comparison. They were immense at the time. Yeah. The the 383 00:21:30,760 --> 00:21:35,077 standard data center that's being built is on the order of 400 megawatts. These 384 00:21:35,160 --> 00:21:39,837 things are plus or minus, and they're about um a half a mile long and about 385 00:21:39,920 --> 00:21:43,557 500 ft wide, right? And they're essentially airflow machines. Take the 386 00:21:43,640 --> 00:21:48,437 air, send the air out, uh cool it in the middle uh using typically air cooling, 387 00:21:48,520 --> 00:21:51,757 and then they have a water system inside to keep the chips 388 00:21:51,840 --> 00:21:54,037 using Nvidia's an example, the um the 389 00:21:54,120 --> 00:21:59,197 chips are water-cooled, and also the HBM3E and now the four memories put out 390 00:21:59,280 --> 00:22:02,597 so much heat that they have to be water-cooled. The chips are 2 kilowatts. 391 00:22:02,680 --> 00:22:05,557 I mean, this is insane. These things will kill you. You want to hear 392 00:22:05,640 --> 00:22:10,677 something truly astounding and funny in hindsight? When you bought DeepMind, Yeah. 393 00:22:10,760 --> 00:22:12,837 everybody thought it was like 800 million bucks or something. 394 00:22:12,920 --> 00:22:14,917 >> 600 million. 600 million. Bargain. 395 00:22:15,000 --> 00:22:16,837 Everyone thought, why on earth would you 396 00:22:16,920 --> 00:22:22,077 waste 600 million dollars on this zero revenue AI? All it does is play Go. 397 00:22:22,160 --> 00:22:25,197 Yeah. And then years later it came out that 398 00:22:25,280 --> 00:22:29,517 the acquisition paid for itself just by controlling the air conditioning more 399 00:22:29,600 --> 00:22:33,237 efficiently in the data centers. The entire acquisition price was paid off, 400 00:22:33,320 --> 00:22:35,597 and that became the AI that's changing the world today. 401 00:22:35,680 --> 00:22:38,757 >> And that one go The credit of that one actually goes to Larry Page. Larry had 402 00:22:38,840 --> 00:22:42,077 studied AI when he was a Stanford graduate student, and he We always 403 00:22:42,160 --> 00:22:43,997 deferred to him in this thing. And he 404 00:22:44,080 --> 00:22:47,357 said, this is the best team. Um and I 405 00:22:47,440 --> 00:22:49,837 think Elon and Larry competed over it. 406 00:22:49,920 --> 00:22:51,637 There was some complicated kerfuffle 407 00:22:51,720 --> 00:22:56,037 there. Yeah. And so um Jeff Dean, who's 408 00:22:56,120 --> 00:22:59,997 the the chief scientist, went over, and then he and I basically, you know, 409 00:23:00,080 --> 00:23:03,517 finished the deal. And I still remember, you know, it there 410 00:23:03,600 --> 00:23:06,837 in one floor, you know, these sort of British people, 411 00:23:06,920 --> 00:23:09,517 and uh led by a sort of Greek British 412 00:23:09,600 --> 00:23:14,157 person, Demis. Yeah. And they were smart, but Google's full of other smart 413 00:23:14,240 --> 00:23:18,317 people. To give you an example, in 2016 414 00:23:18,400 --> 00:23:22,477 um so Demis announces that we're going to win the game of Go. I figured, well, 415 00:23:22,560 --> 00:23:25,637 And by the way, they're at this point they're a separate group. We've We've 416 00:23:25,720 --> 00:23:28,357 let them alone cuz they have to grow and figure out what they're doing and all 417 00:23:28,440 --> 00:23:31,717 that. Okay. And this is the patience of capital. We could let them do that. We 418 00:23:31,800 --> 00:23:34,157 didn't require that they do anything. 419 00:23:34,240 --> 00:23:36,997 And so he goes, um I'm going to go, and I said, well, 420 00:23:37,080 --> 00:23:38,757 I'm going to come, too. 421 00:23:38,840 --> 00:23:44,237 And I said, okay. So, I fly to Korea, and it's all one floor, and I go and I 422 00:23:44,320 --> 00:23:48,157 meet the team that's been winning the game. And of course, these all of these 423 00:23:48,240 --> 00:23:50,797 Koreans They're all very excited about this cuz they know they're going to beat 424 00:23:50,880 --> 00:23:55,437 the computer. >> Mhm. And so, the the Koreans are in one room, 425 00:23:55,520 --> 00:23:58,397 I'm in another. I go to the Korean room, and they're all saying, you know, we're 426 00:23:58,480 --> 00:24:02,037 going to beat beat the crap out of this Google Google Google group. Then I go 427 00:24:02,120 --> 00:24:06,717 into the Google room, and it's very quiet, and there's a monitor. 428 00:24:06,800 --> 00:24:10,997 And there's an basically what I now understand is an RL prediction mechanism 429 00:24:11,080 --> 00:24:13,117 of whether we're winning or not. Yeah, 430 00:24:13,200 --> 00:24:17,037 cool. Okay? And it starts at 50/50. 431 00:24:17,120 --> 00:24:18,917 So, I watch I go listen to the Koreans 432 00:24:19,000 --> 00:24:24,397 talk for a while, and then I go watch the screen, and it goes to 51%. 433 00:24:24,480 --> 00:24:27,957 Okay? And then it goes to 52%. 434 00:24:28,040 --> 00:24:32,237 And then David, who is the guy who's the architect, says, 435 00:24:32,320 --> 00:24:35,647 well, we just planned for it to get to infinity. 436 00:24:35,730 --> 00:24:41,517 >> [laughter] >> Okay. So, basically, it's a it's the abundance theory. So, it's just 437 00:24:41,600 --> 00:24:44,357 And >> [laughter] >> And everyone is All the humans are crushed. 438 00:24:44,440 --> 00:24:47,437 >> Yeah, they're all crying. The DeepMind people say, Yeah, yeah. 439 00:24:47,520 --> 00:24:50,837 They were They were supposed to. Yeah, yeah. [laughter] Welcome. And And then I 440 00:24:50,920 --> 00:24:54,517 understood the genius of the DeepMind people. And you can see this today with 441 00:24:54,600 --> 00:24:57,157 Gemini. Gemini 3 442 00:24:57,240 --> 00:25:00,357 is probably the broadest of the um of of 443 00:25:00,440 --> 00:25:04,757 the non-Chinese systems in terms of its depth. And because it's multilingual, 444 00:25:04,840 --> 00:25:07,717 multimodal, and so forth. >> so much so many moments in your life are 445 00:25:07,800 --> 00:25:11,597 just turning points in history, and I don't know if you realize them in in the 446 00:25:11,680 --> 00:25:15,517 moment, but that was one of the last moments where, you know, we humans used 447 00:25:15,600 --> 00:25:18,517 to look for challenges where the computer could try to catch up, like 448 00:25:18,600 --> 00:25:20,357 chess. And then I think Go was the end 449 00:25:20,440 --> 00:25:22,917 point. Go And And we knew that, by the 450 00:25:23,000 --> 00:25:28,957 way. We understood that Go the game of Go was sort of incomputable by normal 451 00:25:29,040 --> 00:25:31,317 algorithms. >> Yeah. There was lots of math that said 452 00:25:31,400 --> 00:25:34,677 you couldn't solve it. And so it was the combina- They They came up with a 453 00:25:34,760 --> 00:25:39,237 two-tree model with two different RL trees. So, one of these other thing I 454 00:25:39,320 --> 00:25:43,277 learned about these these things is it's not just write me a Go program that will 455 00:25:43,360 --> 00:25:46,157 win the game. You actually have to understand the game 456 00:25:46,240 --> 00:25:50,277 and so forth. So, they took the same team. They They got bored with Go having won. 457 00:25:50,360 --> 00:25:53,037 So, then they took the same team and had them work on protein folding. 458 00:25:53,120 --> 00:25:56,757 >> Yeah. And in protein folding, they took a protein whole bunch of protein 459 00:25:56,840 --> 00:25:59,557 scientists, which I know Can we just think about the genius of that? Like 460 00:25:59,640 --> 00:26:03,677 nobody would ever connect the game of Go >> Demis. solving all of biology. But But 461 00:26:03,760 --> 00:26:06,637 Demis Demis had always wanted to work on this. Larry and Sergey were very 462 00:26:06,720 --> 00:26:10,677 interested in this, and protein folding is the perfect problem because you had 463 00:26:10,760 --> 00:26:13,797 um you got a constant >> end point. You got a defined end point. 464 00:26:13,880 --> 00:26:15,437 So, what happens is people get excited 465 00:26:15,520 --> 00:26:18,717 about AI, but you need to have a 466 00:26:18,800 --> 00:26:22,637 validation function because these things don't have common sense yet. Yeah. 467 00:26:22,720 --> 00:26:25,557 Right? You have to show them what it's got. Ultimately, they produced this 468 00:26:25,640 --> 00:26:27,077 thing called AlphaZero, which 469 00:26:27,160 --> 00:26:32,797 essentially self-learns. Let's talk about data centers in space. Um I'm in 470 00:26:32,880 --> 00:26:36,677 favor of it. >> Eight Eight, nine months ago, no one was discussing this. 471 00:26:36,760 --> 00:26:40,717 >> I mean, all of a sudden >> know why I'm in favor of them? I do, but you [laughter] 472 00:26:40,800 --> 00:26:42,997 you can mention it as you wish. 473 00:26:43,080 --> 00:26:46,757 Um but all of a sudden, everybody's talking about them. 474 00:26:46,840 --> 00:26:52,277 What are your thoughts? Um I'm part owner of a rocket company, and we 475 00:26:52,360 --> 00:26:55,597 need >> Which I love. I love having you come into the space community. It's 476 00:26:55,680 --> 00:26:57,117 >> You understand this far, far better than I. 477 00:26:57,200 --> 00:27:01,077 >> It's It's uh And I've been Rocket science is named that for a reason. Um 478 00:27:01,160 --> 00:27:03,197 rocket science is really, really hard. 479 00:27:03,280 --> 00:27:06,117 Um what I've And I don't know that much about rockets, although I certainly know 480 00:27:06,200 --> 00:27:09,517 how to manage tech people. Um but But I 481 00:27:09,600 --> 00:27:13,997 I I think that the opportunity is is large and interesting. There are 482 00:27:14,080 --> 00:27:17,677 challenges. There's an issue of getting heat off of Again, you know this much 483 00:27:17,760 --> 00:27:21,077 better than I do. Getting heat off of it because you don't have oxygen, and you 484 00:27:21,160 --> 00:27:24,317 also have radiation issues. Yeah. Those have to get addressed. 485 00:27:24,400 --> 00:27:28,797 >> But it makes the business plan for every rocket company that's large enough, 486 00:27:28,880 --> 00:27:29,957 right? The small guys aren't going to 487 00:27:30,040 --> 00:27:34,077 launch it, but Relativity Space and uh I 488 00:27:34,160 --> 00:27:37,357 and you know, Blue Origin and SpaceX, I mean, 489 00:27:37,440 --> 00:27:40,517 I you know, Elon's predictions were, you 490 00:27:40,600 --> 00:27:44,877 know, I think a launch per hour to populate the constellation he wanted. 491 00:27:44,960 --> 00:27:49,797 Um but do you think technically we're going to get there? 492 00:27:49,880 --> 00:27:51,597 Well, the technology is understood. 493 00:27:51,680 --> 00:27:54,597 What's interesting about Technically for the for the data centers in space for 494 00:27:54,680 --> 00:27:57,957 heat dissipation Yeah, that That technology is understood. Uh to me, it's 495 00:27:58,040 --> 00:28:01,637 a business question. Um and the question is what where where 496 00:28:01,720 --> 00:28:04,517 should the data center be? Should it be in space with these other issues, but it 497 00:28:04,600 --> 00:28:08,237 has other benefits, including infinite power and so forth, Yeah. versus on the 498 00:28:08,320 --> 00:28:12,397 ground where you have fiber, and you know, it's not shaking too much. Yeah. I 499 00:28:12,480 --> 00:28:17,677 mean, the energy argument says space wins by far, and I think the the cooling 500 00:28:17,760 --> 00:28:21,237 is a very big challenge, but I think it's largely figured out now. 501 00:28:21,320 --> 00:28:24,917 But then the politics of space You know, one of the turning points in 502 00:28:25,000 --> 00:28:27,677 AI history also is you getting in front of Congress and saying, hey, we need to 503 00:28:27,760 --> 00:28:32,877 find almost 100 gigawatts. And at the time, it seemed outrageous. 504 00:28:32,960 --> 00:28:35,957 And now, of course, it's mainstream, and it looks like, you know, the crazy 505 00:28:36,040 --> 00:28:39,197 investors are solving the problem. You They're unleashing of the money is 506 00:28:39,280 --> 00:28:42,477 happening. >> American capitalism. Tech industry, I 507 00:28:42,560 --> 00:28:44,277 mean, we have another set of problems. 508 00:28:44,360 --> 00:28:46,757 Well, if the next frontier is space, then but there's no there's no 509 00:28:46,840 --> 00:28:50,277 investment community in space, but there's also no military 510 00:28:50,360 --> 00:28:53,797 jurisdiction in space, or maybe there is. I I would have never thought my 511 00:28:53,880 --> 00:28:57,797 childhood dreams of going to the moon and Mars and would be fueled by data 512 00:28:57,880 --> 00:29:02,557 centers. Um Never. Well, mine can't carry you at the 513 00:29:02,640 --> 00:29:04,597 moment, but it can in the future, maybe. 514 00:29:04,680 --> 00:29:09,597 Yeah. Yeah, all the way. I I read an a Time op-ed piece you wrote 515 00:29:09,680 --> 00:29:11,357 last night. Um 516 00:29:11,440 --> 00:29:16,717 China can dominate the physical AI future, or uh What did you Can you 517 00:29:16,800 --> 00:29:19,117 summarize that for us? It was an important conversation. 518 00:29:19,200 --> 00:29:24,037 >> the geopolitical context, the And I've said this many times, I'll 519 00:29:24,120 --> 00:29:27,797 say it again. The American competitor, not enemy, but 520 00:29:27,880 --> 00:29:31,517 competitor, is China. I I think it's By the way, I think it's 521 00:29:31,600 --> 00:29:33,397 an important distinction for you to make, so thank you. 522 00:29:33,480 --> 00:29:39,117 >> Not Not enemy, competitor. Um and how to understand them as a 523 00:29:39,200 --> 00:29:42,437 competitor. They have lots of money. They're very, 524 00:29:42,520 --> 00:29:45,317 very smart. Their work ethic is equal or 525 00:29:45,400 --> 00:29:47,917 stronger than ours, and they dominate 526 00:29:48,000 --> 00:29:50,757 key industries. And I don't know Right? 527 00:29:50,840 --> 00:29:54,077 With respect to robotics, we somehow 528 00:29:54,160 --> 00:29:56,917 decided it was okay for them to dominate 529 00:29:57,000 --> 00:30:01,317 the electric vehicle industry. This was an error. To be very clear, it's an 530 00:30:01,400 --> 00:30:05,197 error. Why do you not understand it's an error? Because we don't allow their cars 531 00:30:05,280 --> 00:30:10,237 in. Spend some time outside of this country in Chinese cars, trust me. 532 00:30:10,320 --> 00:30:11,797 They are real competitors. They've done 533 00:30:11,880 --> 00:30:14,997 a great job. As I understand it, 534 00:30:15,080 --> 00:30:19,797 China is capable of vertical integration and building these gigafactories at a 535 00:30:19,880 --> 00:30:24,277 scale that we can't for all sorts of reasons. That's got to get addressed. 536 00:30:24,360 --> 00:30:27,997 So, if you want to compete and I want to compete and win with China and the and 537 00:30:28,080 --> 00:30:31,557 the and I compete not enemy. I want us to have the same kind of 538 00:30:31,640 --> 00:30:36,317 system. So, in robotics, it turns out robotics, 539 00:30:36,400 --> 00:30:39,597 you can understand robots as essentially actuators. These little stepper motors, 540 00:30:39,680 --> 00:30:44,077 click click click click click click and a brain. Ignoring the appearance and the 541 00:30:44,160 --> 00:30:46,797 Google, you know, eyes and all that kind of stuff. 542 00:30:46,880 --> 00:30:49,197 It turns out that the industry of the 543 00:30:49,280 --> 00:30:51,557 electric vehicle produces the same kind 544 00:30:51,640 --> 00:30:54,957 of motors and the same kind of systems. 545 00:30:55,040 --> 00:30:57,717 They have an expertise that we don't. 546 00:30:57,800 --> 00:31:02,197 My own view is that at least for very low cost, 547 00:31:02,280 --> 00:31:06,237 China is going to win that. And that was what I was trying to say in that piece. 548 00:31:06,320 --> 00:31:10,477 And I worry about that. Now, today these are not particularly useful. You know, 549 00:31:10,560 --> 00:31:13,277 they're they're fun toys. They're a replacement for the dog if you get mad 550 00:31:13,360 --> 00:31:15,277 at your dog. Sorry, I love I love dogs. 551 00:31:15,360 --> 00:31:18,597 But you get the idea. So, 552 00:31:18,680 --> 00:31:21,277 so we we we need to address this. 553 00:31:21,360 --> 00:31:25,597 But at the moment, it sure looks to me like the robotic hardware 554 00:31:25,680 --> 00:31:28,277 of China is the winner at the low end. 555 00:31:28,360 --> 00:31:30,797 I'm not talking about the high end. I'm not talking about the expensive stuff. 556 00:31:30,880 --> 00:31:32,757 I'm not talking about industrial robots. 557 00:31:32,840 --> 00:31:34,877 And if you're confused, watch the 558 00:31:34,960 --> 00:31:39,837 Unitree robot dance with the humans. 559 00:31:39,920 --> 00:31:41,437 Yeah. That came out about a month ago. 560 00:31:41,520 --> 00:31:44,637 >> Unitree here in the tech hub and the co-founder will be on stage with us 561 00:31:44,720 --> 00:31:46,717 later today. Pay Pay attention to them. 562 00:31:46,800 --> 00:31:48,757 They're they're very very impressive. 563 00:31:48,840 --> 00:31:51,917 And I spent some time with them last time I was in China. 564 00:31:52,000 --> 00:31:53,997 And they're one of many. And the way 565 00:31:54,080 --> 00:31:56,837 China works is that they have brutal 566 00:31:56,920 --> 00:32:01,357 competition. Brutal brutal It's like unbelievable. I was talking to my 567 00:32:01,440 --> 00:32:05,677 friend, we teach at Stanford, and he said, you know, in China, we don't have 568 00:32:05,760 --> 00:32:08,437 the board dinner. We have a 2-hour meeting. We get back to work. And 569 00:32:08,520 --> 00:32:11,117 there's no preamble. We're not, you know, we don't say hello, how are you? 570 00:32:11,200 --> 00:32:12,317 How's the family? That kind of stuff. 571 00:32:12,400 --> 00:32:17,877 We're boom boom. Right? It's just cultural. Yeah. And the 572 00:32:17,960 --> 00:32:21,237 work the the work ethic, the precision, and 573 00:32:21,320 --> 00:32:25,277 the scale that is possible in China is a real competitor. 574 00:32:25,360 --> 00:32:28,237 I don't want to lose the robotic 575 00:32:28,320 --> 00:32:31,077 revolution in my view the way we lost 576 00:32:31,160 --> 00:32:35,757 the electric vehicle revolution at least on the low end. Interesting. Cuz I had 577 00:32:35,840 --> 00:32:39,077 the Chinese model very much has a well-built-out supply chain with many 578 00:32:39,160 --> 00:32:43,077 vendors in the loop. But we we've been on a worldwide tour of all of the 579 00:32:43,160 --> 00:32:47,037 humanoid robotics companies. And uh just like coincidence, I guess. But when you 580 00:32:47,120 --> 00:32:50,157 look at the gigafactory and look at Elon's vertical integration and also 581 00:32:50,240 --> 00:32:53,517 Brad Adcock, it figures same thing. It's all vertically integrated. Why? Well, 582 00:32:53,600 --> 00:32:58,357 cuz there's no vendor. Yeah. I I have no choice. I have So, so it appears that to 583 00:32:58,440 --> 00:33:01,997 get to again, we're in this is the abundance group. This is the abundance 584 00:33:02,080 --> 00:33:06,117 club. The way you get to abundance is you drive prices down and you get 585 00:33:06,200 --> 00:33:10,077 vertically integrated. And Elon in our country pioneered that to his credit. 586 00:33:10,160 --> 00:33:13,797 Yeah. Right? You know, the old joke about Google was we would build anything 587 00:33:13,880 --> 00:33:18,237 including the buildings. Well, Elon is actually doing that. Yeah. You 588 00:33:18,320 --> 00:33:20,597 know, he And why? Not because he's 589 00:33:20,680 --> 00:33:25,917 insane, but because he actually That's how you drive cost down. Yeah. He He 590 00:33:26,000 --> 00:33:27,517 truly believes, and I'd love to get your 591 00:33:27,600 --> 00:33:29,637 your take on this. He truly believes 592 00:33:29,720 --> 00:33:32,957 that the the robot building the robot is 593 00:33:33,040 --> 00:33:37,157 imminent. And I didn't get it until we toured the gigafactory and I realized 594 00:33:37,240 --> 00:33:40,637 that almost all of it's automated already. It's just the last piece is the 595 00:33:40,720 --> 00:33:44,677 human controlling a few knobs. And that, you know, the humanoid robot can 596 00:33:44,760 --> 00:33:46,717 actually do. >> Let me define the boundary cuz it's 597 00:33:46,800 --> 00:33:51,437 important. Let's use batteries, for example. Batteries are predictable, 598 00:33:51,520 --> 00:33:56,237 straightforward manufacturing processes at huge scale. Right? Things like that 599 00:33:56,320 --> 00:33:59,357 will have gigafactories. One of the questions and I of course 600 00:33:59,440 --> 00:34:04,117 used LLMs to do my deep research as a new person in the rocket company was how 601 00:34:04,200 --> 00:34:08,517 much of the human labor to build a rocket could be replaced by robots. 602 00:34:08,600 --> 00:34:13,557 The current limit, which of course will change, is in our company, we have these 603 00:34:13,640 --> 00:34:15,117 extraordinarily 604 00:34:15,200 --> 00:34:18,917 talented essentially assembly people. 605 00:34:19,000 --> 00:34:22,637 And they're they're they're more than welders and they're more than, you know, 606 00:34:22,720 --> 00:34:28,277 mechanics. They understand precisely how the little the tubes and so forth and so 607 00:34:28,360 --> 00:34:30,917 on go together and they use precise 608 00:34:31,000 --> 00:34:33,637 tolerances. That kind of assembly is 609 00:34:33,720 --> 00:34:39,397 beyond current robots. I'm sure it will eventually show up, but 610 00:34:39,480 --> 00:34:42,477 not for a long time. >> People don't realize the majority cost 611 00:34:42,560 --> 00:34:46,397 of a rocket is labor. Exactly. It's not material. You'll listen to this person. 612 00:34:46,480 --> 00:34:49,717 >> when they get inside the rocket, they understand what they're doing and they 613 00:34:49,800 --> 00:34:53,677 see what's wrong and they use human judgment. We don't have those systems 614 00:34:53,760 --> 00:34:57,517 yet. Yeah. Now, perhaps we will in the future or perhaps this will be one of 615 00:34:57,600 --> 00:34:59,637 the last things to go. 616 00:34:59,720 --> 00:35:02,557 Right? But at the moment, high-skilled 617 00:35:02,640 --> 00:35:05,637 mechanical labor is very important. 618 00:35:05,720 --> 00:35:11,077 Low-skilled labor of any kind gets swept up. Well, so both of those, there's AI 619 00:35:11,160 --> 00:35:14,237 self-improvement and then there's robotics building robots building robots 620 00:35:14,320 --> 00:35:16,237 self-improvement. Both of those loops. 621 00:35:16,320 --> 00:35:19,957 And I I we we talk a lot about closed loops being 622 00:35:20,040 --> 00:35:21,437 If I can interrupt you. Yeah, please. 623 00:35:21,520 --> 00:35:24,357 The term I like to use is learning loops. Learning loops. In a business, 624 00:35:24,440 --> 00:35:27,437 try to figure out all the different learning loops and then try to 625 00:35:27,520 --> 00:35:30,677 accelerate the learning. Fastest learner wins. 626 00:35:30,760 --> 00:35:33,037 Sorry. Well, I'm going to lead the witness here a little bit. But if you if 627 00:35:33,120 --> 00:35:37,157 you asked Daniela Rus over at CSAIL a year ago, Erik Brynjolfsson at Stanford 628 00:35:37,240 --> 00:35:38,437 a year ago, 629 00:35:38,520 --> 00:35:40,797 do we need one more big breakthrough in 630 00:35:40,880 --> 00:35:44,437 AI core science or will scaling what 631 00:35:44,520 --> 00:35:46,277 we've already got lead to 632 00:35:46,360 --> 00:35:48,797 self-improvement and then that will be 633 00:35:48,880 --> 00:35:51,957 all we need to get to AGI 634 00:35:52,040 --> 00:35:55,957 >> and then self-improvement. That that I spent the last week doing RSI reviews, 635 00:35:56,040 --> 00:35:58,157 recursive self-improvement 636 00:35:58,240 --> 00:36:02,517 reviews. The scientists do not agree on 637 00:36:02,600 --> 00:36:06,717 the exact approach to work yet. So, I think it's too early to know that 638 00:36:06,800 --> 00:36:09,637 question. There's evidence that it will work. 639 00:36:09,720 --> 00:36:12,077 There are tests in the lab that show it. 640 00:36:12,160 --> 00:36:15,797 But they show it in limited cases that are kind of demos. 641 00:36:15,880 --> 00:36:17,677 Right? Real recursive self-improvement 642 00:36:17,760 --> 00:36:22,997 is the following. Start now, learn everything, discover things, and tell me 643 00:36:23,080 --> 00:36:24,597 what you learned. Yeah. 644 00:36:24,680 --> 00:36:27,837 That query doesn't work yet. 645 00:36:27,920 --> 00:36:32,037 We're seeing all of the frontier labs sort of constantly leapfrog each other, 646 00:36:32,120 --> 00:36:34,397 right? I mean, literally every week it's the new model. 647 00:36:34,480 --> 00:36:36,797 >> Unbelievable. And it's it's extraordinary. 648 00:36:36,880 --> 00:36:38,717 And do you imagine they're all 649 00:36:38,800 --> 00:36:44,680 converging towards the same endpoint or is anyone going to pull ahead? 650 00:36:45,200 --> 00:36:46,757 So, if you go back to this question of 651 00:36:46,840 --> 00:36:49,477 capital, how many how much room is there 652 00:36:49,560 --> 00:36:54,277 in the world for these companies? How many can there be? 653 00:36:54,360 --> 00:36:56,317 I'm going to make some numbers up. 654 00:36:56,400 --> 00:37:00,597 And these are made up. I think there's at least 10 in the world 655 00:37:00,680 --> 00:37:04,157 at this scale. I think there'll be a few in in China. I 656 00:37:04,240 --> 00:37:08,437 think the majority will be in the United States, the usual suspects most likely. 657 00:37:08,520 --> 00:37:12,517 I think there'll be a They They might be one or two in Europe. Mhm. Depending 658 00:37:12,600 --> 00:37:14,517 their electricity costs are a problem. 659 00:37:14,600 --> 00:37:16,997 There might be one in India, right? 660 00:37:17,080 --> 00:37:20,997 There's not going to be one in Russia because of the the war and so forth and 661 00:37:21,080 --> 00:37:22,117 so on. 662 00:37:22,200 --> 00:37:26,277 So, can the world accommodate 10? Yes. 663 00:37:26,360 --> 00:37:30,037 Do they track together? I don't think so. 664 00:37:30,120 --> 00:37:33,397 One of the key things to understand about China is China's approach, which 665 00:37:33,480 --> 00:37:36,637 has produced DCV4, Quinn, Kimmy 2, Kimmy 666 00:37:36,720 --> 00:37:41,117 version 3, etc. and more coming, is it's all open 667 00:37:41,200 --> 00:37:45,677 source open weights. They've managed to do this with our chip 668 00:37:45,760 --> 00:37:49,757 limitations against them, which annoy them no end. Shows you how clever they 669 00:37:49,840 --> 00:37:54,517 are. They're also the Chinese strategy is a bit different. It's less central 670 00:37:54,600 --> 00:37:58,117 computing and much more edge computing, which has to do with enveloping their 671 00:37:58,200 --> 00:38:02,917 Chinese customers with AI around them all all the time. We're much more AGI 672 00:38:03,000 --> 00:38:06,117 ASI centered, which is fine. 673 00:38:06,200 --> 00:38:09,357 So, the the patterns are diverging. 674 00:38:09,440 --> 00:38:14,277 Within the companies, it's it's it's a jumble now. But 675 00:38:14,360 --> 00:38:18,037 fundamentally, Microsoft and Google have these large cash flow streams from 676 00:38:18,120 --> 00:38:20,677 enterprise, so they can fund that. Yeah. 677 00:38:20,760 --> 00:38:24,437 Anthropic has done a fantastic job of raising money from those other companies 678 00:38:24,520 --> 00:38:26,157 to and they use, as you know, Google 679 00:38:26,240 --> 00:38:29,237 TPUs. And they've become the sort of 680 00:38:29,320 --> 00:38:33,797 leading player in this sort of Claude API within the enterprise agent 681 00:38:33,880 --> 00:38:38,917 system. And and OpenAI is now shifting a bit of its strategy to include the the new 682 00:38:39,000 --> 00:38:42,397 things it's doing. I don't think we can predict. 683 00:38:42,480 --> 00:38:46,597 But the key thing to understand is that they need so much money. 684 00:38:46,680 --> 00:38:50,397 Right? I mean, you look at what Sam is trying to raise. They need so much money 685 00:38:50,480 --> 00:38:54,357 that they're forced into these situations where they have to win those 686 00:38:54,440 --> 00:38:55,677 battles. And they're busy winning them. 687 00:38:55,760 --> 00:38:59,877 I mean, this is all good. I think it's the it's We In a year, we'll know better 688 00:38:59,960 --> 00:39:03,517 the answer to your question. I two questions I want to close this out with 689 00:39:03,600 --> 00:39:05,797 that I think are important. Um 690 00:39:05,880 --> 00:39:08,117 You made a a statement a couple of 691 00:39:08,200 --> 00:39:11,237 times, once in our podcast, once uh 692 00:39:11,320 --> 00:39:13,157 elsewhere that 693 00:39:13,240 --> 00:39:16,237 regarding AI safety, the world may need 694 00:39:16,320 --> 00:39:19,517 to have a modest Chernobyl-like death 695 00:39:19,600 --> 00:39:23,397 event in order for us to wake up. 696 00:39:23,480 --> 00:39:26,037 Uh do you still believe that's the case? 697 00:39:26,120 --> 00:39:27,677 Yeah, and by the way, I'm not endorsing 698 00:39:27,760 --> 00:39:33,760 that. I'm I'm doing it as a descriptive, not as a as a proscriptive. 699 00:39:35,080 --> 00:39:37,437 What are the real dangers of this? There 700 00:39:37,520 --> 00:39:40,437 are biological dangers. I mean, there's 701 00:39:40,520 --> 00:39:44,557 obviously the dangers of of kids and democracies and so forth. But 702 00:39:44,640 --> 00:39:46,957 like let's think about like a biological 703 00:39:47,040 --> 00:39:49,917 attack, a nuclear attack that's spawned 704 00:39:50,000 --> 00:39:52,117 by these things. 705 00:39:52,200 --> 00:39:55,757 It may take such a tragedy, hopefully a 706 00:39:55,840 --> 00:39:59,357 small one, to awaken the world to understand that 707 00:39:59,440 --> 00:40:03,517 these things are they do have negative power. 708 00:40:03,600 --> 00:40:06,877 And so I can imagine, I'm making this up, that there's some bad thing that 709 00:40:06,960 --> 00:40:10,277 happened and then all of the leaders, 710 00:40:10,360 --> 00:40:14,797 China, the US, we all have a meeting and we basically say like what are we 711 00:40:14,880 --> 00:40:18,597 going to do? You know, we're in brutal competition, we hate 712 00:40:18,680 --> 00:40:20,957 each other, you know, I don't like you, you know, you don't speak the same 713 00:40:21,040 --> 00:40:22,877 language, blah blah blah, but we are all 714 00:40:22,960 --> 00:40:26,237 in it together, right? Over this issue. 715 00:40:26,320 --> 00:40:30,920 My sense is that will happen, but I don't know when. 716 00:40:31,320 --> 00:40:34,997 We had we had a congressperson on this stage last year and one of the crowd 717 00:40:35,080 --> 00:40:38,717 asked, how much time do you spend talking about AI in Congress? 718 00:40:38,800 --> 00:40:42,557 And he said, well, it's definitely way less than 1%. Yeah. And so without that 719 00:40:42,640 --> 00:40:45,997 wake-up call, I don't see how you have >> I mean governments, I spend lots of time 720 00:40:46,080 --> 00:40:49,037 with governments. Governments are super busy, right? And they're and they're 721 00:40:49,120 --> 00:40:52,717 driven by political things, they're driven by by 722 00:40:52,800 --> 00:40:56,477 in at least in democracies by by political sentiment. 723 00:40:56,560 --> 00:40:59,237 I'd like us as a nation to focus on the 724 00:40:59,320 --> 00:41:02,517 following. I want to win the AI race. I 725 00:41:02,600 --> 00:41:06,517 want us to do whatever it takes to do that. This government is doing a very 726 00:41:06,600 --> 00:41:10,437 good job of making energy permitting more accessible. 727 00:41:10,520 --> 00:41:14,877 The rate at which data centers are getting built is now accelerated, 728 00:41:14,960 --> 00:41:19,557 solving the grid problems. I also want lots of immigrants in our country 729 00:41:19,640 --> 00:41:23,437 because those immigrants, at least the high stakes high skills immigration, is 730 00:41:23,520 --> 00:41:25,877 what we need because we need the the 731 00:41:25,960 --> 00:41:28,157 smartest people in the world on our side 732 00:41:28,240 --> 00:41:30,797 to build these systems. This is a unique 733 00:41:30,880 --> 00:41:36,940 moment in history, right? >> [applause] 734 00:41:37,600 --> 00:41:40,317 >> I would take us home on a positive note. 735 00:41:40,400 --> 00:41:43,477 So, you know, you said we're going to get to 736 00:41:43,560 --> 00:41:48,597 ASI at some point, whether it's 2 years out, 5 years out, some place in this 737 00:41:48,680 --> 00:41:52,637 next decade. So, the question is, what steps can we 738 00:41:52,720 --> 00:41:56,917 take to steer artificial superintelligence towards 739 00:41:57,000 --> 00:41:58,237 abundance, 740 00:41:58,320 --> 00:42:02,437 towards uplifting and in alignment with 741 00:42:02,520 --> 00:42:05,717 humanity? You know, to make this abundance thesis 742 00:42:05,800 --> 00:42:09,997 materialize, what's your advice to us, 743 00:42:10,080 --> 00:42:12,637 companies, governments? 744 00:42:12,720 --> 00:42:15,037 There's an over-reliance in our society 745 00:42:15,120 --> 00:42:17,557 on people like me 746 00:42:17,640 --> 00:42:20,477 to work on this. 747 00:42:20,560 --> 00:42:23,877 Why don't we have the smartest people in 748 00:42:23,960 --> 00:42:26,077 politics, history, 749 00:42:26,160 --> 00:42:28,957 human psychology, governance, ethics 750 00:42:29,040 --> 00:42:33,277 working together to make sure this stuff stays in human values and human 751 00:42:33,360 --> 00:42:37,597 alignment? I want the system that we build in 752 00:42:37,680 --> 00:42:42,037 America to reflect American values. The values of freedom and freedom of speech 753 00:42:42,120 --> 00:42:45,517 freedom of association, all those things you learned in college and excuse me, in 754 00:42:45,600 --> 00:42:47,477 elementary school and and high school. 755 00:42:47,560 --> 00:42:49,597 They're still important to our nation. 756 00:42:49,680 --> 00:42:52,997 They're the enablers for the next generation of our children and 757 00:42:53,080 --> 00:42:56,437 grandchildren. I I desperately want that 758 00:42:56,520 --> 00:43:00,157 and I don't want America to ever get on the wrong side of that battle. 759 00:43:00,240 --> 00:43:02,597 There's lots of people working on this. 760 00:43:02,680 --> 00:43:06,117 Lots of people understand the technical details. I happen to run an informal 761 00:43:06,200 --> 00:43:08,317 group that discusses this every week. 762 00:43:08,400 --> 00:43:12,677 So, it's possible, but it requires political 763 00:43:12,760 --> 00:43:15,397 will and it requires an understanding that 764 00:43:15,480 --> 00:43:19,917 this can be done without screwing up the genius of America, right? In other 765 00:43:20,000 --> 00:43:24,317 words, I'm not suggesting slowing anything down. I'm you said it so well, 766 00:43:24,400 --> 00:43:28,517 shaping it. Making sure we don't cross lines. Like I 767 00:43:28,600 --> 00:43:32,597 already mentioned the teenage sorry, the underage kids problem. That's a line we 768 00:43:32,680 --> 00:43:37,520 can't cross. We got to solve that problem. There are others.