1 00:00:03,000 --> 00:00:05,880 Hello everyone. 2 00:00:07,760 --> 00:00:11,997 Well, welcome to lunch. Not for you and me, but welcome to lunch. 3 00:00:12,080 --> 00:00:14,037 Um 4 00:00:14,120 --> 00:00:17,437 Uh I think we've had a great day so far. 5 00:00:17,520 --> 00:00:21,037 And uh and I'm honored to have with me 6 00:00:21,120 --> 00:00:24,397 uh Sam Altman uh joining us for our 7 00:00:24,480 --> 00:00:27,077 discussion about uh 8 00:00:27,160 --> 00:00:29,877 AI and and where we're going from here. 9 00:00:29,960 --> 00:00:34,037 So, um So, thanks for thanks for joining us. 10 00:00:34,120 --> 00:00:37,717 You have really the greatest countries in the world all 11 00:00:37,800 --> 00:00:42,237 around you. So, would you mind giving us a little history like 12 00:00:42,320 --> 00:00:46,917 how do how do you go from being Sam Altman to being the head of AI and 13 00:00:47,000 --> 00:00:51,797 really at the frontier like sort of walk us through how you got from there to here. 14 00:00:51,880 --> 00:00:54,717 >> First of all, thank you very much for having me. Um as you said, this is 15 00:00:54,800 --> 00:00:57,997 greatest countries in the world and it's it's a great honor to get to address all 16 00:00:58,080 --> 00:01:01,317 of you at the same time. Um 17 00:01:01,400 --> 00:01:05,997 I I was always an AI nerd. I was like a real sci-fi nerd as a kid. I thought AI 18 00:01:06,080 --> 00:01:09,277 was the coolest thing. I always wanted to work on it. Um 19 00:01:09,360 --> 00:01:10,997 I never like really thought I'd actually 20 00:01:11,080 --> 00:01:16,717 get to do it. I kind of can't believe it's happened to me, but I I always 21 00:01:16,800 --> 00:01:19,597 I always thought it was awesome. I went to school to study it. At the time 22 00:01:19,680 --> 00:01:23,477 nothing was working and I kind of gave up on it and I went on the side quest of 23 00:01:23,560 --> 00:01:25,837 entrepreneurship, which was also great. 24 00:01:25,920 --> 00:01:29,757 Um and then in 20 12 25 00:01:29,840 --> 00:01:32,877 um everything changed. None of us paid 26 00:01:32,960 --> 00:01:37,317 enough attention, but there was a a real watershed moment where all of a sudden 27 00:01:37,400 --> 00:01:41,517 deep learning started to work and it got better with more compute. And I paid a 28 00:01:41,600 --> 00:01:43,917 lot of attention to that for the next couple of years. 29 00:01:44,000 --> 00:01:48,637 And by the time 2015 rolled around, I was like, "Hey, this is really 30 00:01:48,720 --> 00:01:51,117 this is going to happen. Like we got to take it we got to try. This could be 31 00:01:51,200 --> 00:01:54,557 like the most important thing ever." Uh and so we thought for a while about what 32 00:01:54,640 --> 00:01:58,317 to do and started OpenAI. Uh and 33 00:01:58,400 --> 00:02:00,397 >> So, what year did you start OpenAI? 34 00:02:00,480 --> 00:02:03,357 >> Uh the the very end of 2015. Um we 35 00:02:03,440 --> 00:02:08,240 announced it, and then the first day of work was like January of 2016. 36 00:02:08,520 --> 00:02:11,437 >> So, all right. So, you start in uh in in 37 00:02:11,520 --> 00:02:14,357 2016, and and what are your like what's 38 00:02:14,440 --> 00:02:19,157 the big goal that you start with, and then how does that evolve? 39 00:02:19,240 --> 00:02:20,677 >> So, we knew that there was this magic 40 00:02:20,760 --> 00:02:22,597 piece of technology. We knew that we 41 00:02:22,680 --> 00:02:26,597 could apply compute 42 00:02:26,680 --> 00:02:28,917 to a problem and to learn it. And we 43 00:02:29,000 --> 00:02:33,997 knew that the more compute we used, the smarter and more capable the system got. 44 00:02:34,080 --> 00:02:39,717 We didn't really know anything else. We didn't know how we were going to um 45 00:02:39,800 --> 00:02:41,917 what kind of system we were going to build. This is way before the idea of 46 00:02:42,000 --> 00:02:47,237 language models. We were trying to train a robot hand to do a Rubik's Cube. Um we 47 00:02:47,320 --> 00:02:49,677 were trying to you know, show that we could play video 48 00:02:49,760 --> 00:02:53,517 games. And we were really developing the algorithms and trying to understand how 49 00:02:53,600 --> 00:02:55,797 this amazing property of the universe 50 00:02:55,880 --> 00:02:56,837 worked. 51 00:02:56,920 --> 00:03:00,957 And and then not until about 20 18 52 00:03:01,040 --> 00:03:05,277 did we start making progress on unsupervised learning, and then the 53 00:03:05,360 --> 00:03:07,837 first GPT-1 model. 54 00:03:07,920 --> 00:03:11,357 And not until 2020 did we train GPT-3, 55 00:03:11,440 --> 00:03:15,397 which was the first model that I think actually was over some threshold of of 56 00:03:15,480 --> 00:03:19,637 usability. Even then, most people in the world looked at this and said, 57 00:03:19,720 --> 00:03:24,277 "You know, okay, the model can like generate a coherent sentence." That was 58 00:03:24,360 --> 00:03:27,357 really the limit at the time. Kind of impressive, but like what is this going 59 00:03:27,440 --> 00:03:31,997 to what is this going to mean? And why is why is this going to like matter to the 60 00:03:32,080 --> 00:03:34,317 economy or society or anything else? 61 00:03:34,400 --> 00:03:38,837 Um but we had discovered by this point these scaling laws, which said, "Well, 62 00:03:38,920 --> 00:03:42,877 if we could, you know, apply like a trillion dollars of compute to this 63 00:03:42,960 --> 00:03:45,797 idea, something quite amazing would happen." 64 00:03:45,880 --> 00:03:48,797 And people said, "Well, you can't apply a trillion dollars of compute, so it 65 00:03:48,880 --> 00:03:53,877 doesn't matter." And we sort of said, "Why not? Like, let's try. 66 00:03:53,960 --> 00:03:55,797 So, 67 00:03:55,880 --> 00:03:57,597 you've got the idea 68 00:03:57,680 --> 00:04:01,437 and and it's moving forward, but when 69 00:04:01,520 --> 00:04:03,957 when does it come together that it's 70 00:04:04,040 --> 00:04:07,157 really bottling knowledge? 71 00:04:07,240 --> 00:04:09,997 Like, how how does it 72 00:04:10,080 --> 00:04:11,717 >> Different people at OpenAI would give 73 00:04:11,800 --> 00:04:14,237 you different answers to this, but I 74 00:04:14,320 --> 00:04:17,157 would say most the kind of consensus to 75 00:04:17,240 --> 00:04:22,840 the degree there is one was the time of GPT-4. This was like 76 00:04:24,440 --> 00:04:27,957 I believe this was like March of 2023, but I may be wrong on the timeline 77 00:04:28,040 --> 00:04:31,797 there. Um we had that model for maybe 8 months 78 00:04:31,880 --> 00:04:35,717 before it launched, so we had had time internally to really get to know it. And 79 00:04:35,800 --> 00:04:39,277 that was the time where people even some of the people internally had 80 00:04:39,360 --> 00:04:41,517 been skeptics of this approach 81 00:04:41,600 --> 00:04:43,877 would looked at it and said, 82 00:04:43,960 --> 00:04:48,357 "This is going to go very far." Um and 83 00:04:48,440 --> 00:04:52,157 it hasn't been that long since that moment, you know, 3 and 1/2 years now. 84 00:04:52,240 --> 00:04:54,357 Uh and 85 00:04:54,440 --> 00:04:59,277 and the fact that we now have a system 86 00:04:59,360 --> 00:05:02,037 uh and we're launching our new model soon, which is another step forward and 87 00:05:02,120 --> 00:05:06,237 there will be many more steps forward coming soon after that. Um but that it 88 00:05:06,320 --> 00:05:08,357 is so capable that it is discovering new 89 00:05:08,440 --> 00:05:11,077 knowledge, doing science, able to add 90 00:05:11,160 --> 00:05:15,917 huge economic value, create entire pieces of complex software. 91 00:05:16,000 --> 00:05:19,117 Um this is 92 00:05:19,200 --> 00:05:23,197 this is like a very crazy moment. I think it hit us 93 00:05:23,280 --> 00:05:29,160 Yeah, over a period of time, but centered around this GPT-4 period. 94 00:05:29,280 --> 00:05:32,797 So, you understand that 95 00:05:32,880 --> 00:05:35,477 you can sort of capture education. 96 00:05:35,560 --> 00:05:39,917 Right? You capture knowledge. Um how 97 00:05:40,000 --> 00:05:46,320 how should the the G20 economies think about that? 98 00:05:48,680 --> 00:05:52,357 I >> I don't want to overstate things, but I'm like very excited about our new 99 00:05:52,440 --> 00:05:54,477 model, so forgive me if I get a little 100 00:05:54,560 --> 00:05:59,640 bit too enthusiastic here. Um I 101 00:06:01,040 --> 00:06:05,197 I think this, for example, will be the greatest boom in entrepreneurship and 102 00:06:05,280 --> 00:06:07,717 the creation and development of small businesses 103 00:06:07,800 --> 00:06:09,717 um the world has ever seen. 104 00:06:09,800 --> 00:06:12,357 If you have a great idea, but you didn't 105 00:06:12,440 --> 00:06:15,437 have the resources or the ability to get 106 00:06:15,520 --> 00:06:19,997 enough capital together or hire enough experts to work on it, or to figure out 107 00:06:20,080 --> 00:06:23,557 what it's going to take to invent a new piece of science, or manage a complex 108 00:06:23,640 --> 00:06:26,197 supply chain, you can do that 109 00:06:26,280 --> 00:06:29,837 with AI now. Um and I remember after 110 00:06:29,920 --> 00:06:34,197 ChatGPT launched, one of the stories that has stuck with me the most from 111 00:06:34,280 --> 00:06:38,197 that moment was a guy that managed a laundromat. Apparently, laundromats are 112 00:06:38,280 --> 00:06:40,597 usually good businesses. Wouldn't have been obvious to me, 113 00:06:40,680 --> 00:06:42,357 apparently they are. Um but this one was 114 00:06:42,440 --> 00:06:45,477 not going well, and he was using ChatGPT 115 00:06:45,560 --> 00:06:47,717 for everything. Um to make his marketing 116 00:06:47,800 --> 00:06:52,877 materials, to, you know, manage give him advice on legal contracts, to manage 117 00:06:52,960 --> 00:06:56,917 like relationships with vendors. This long list of stuff. He He was really the 118 00:06:57,000 --> 00:07:01,317 first person I met that had figured out how to use AI to 119 00:07:01,400 --> 00:07:03,877 uh effectively run run a business. 120 00:07:03,960 --> 00:07:05,717 And this was amazing to me at the time. 121 00:07:05,800 --> 00:07:08,357 Now, I know you all have heard stories about this a bunch, and it's no longer 122 00:07:08,440 --> 00:07:12,477 surprising, but in 2023, the fact that someone was using 123 00:07:12,560 --> 00:07:16,997 ChatGPT, broken and embarrassing though it was at the time, to let them manage a 124 00:07:17,080 --> 00:07:20,957 business that otherwise they couldn't have done, was was like quite 125 00:07:21,040 --> 00:07:25,637 remarkable. Um In 2025, we saw a real acceleration of 126 00:07:25,720 --> 00:07:29,517 this with coding agents and and white-collar work agents. 127 00:07:29,600 --> 00:07:33,037 Um I think we'll see a third thing soon, where we'll have these long-term 128 00:07:33,120 --> 00:07:34,677 persistent agents that can work like a 129 00:07:34,760 --> 00:07:40,680 real virtual collaborator, constantly helping, but um 130 00:07:40,840 --> 00:07:45,557 it it's it's a it's amazing to me what is happening with 131 00:07:45,640 --> 00:07:49,317 the boom in entrepreneurship and business creation 132 00:07:49,400 --> 00:07:53,197 because of AI. If I were a country, um 133 00:07:53,280 --> 00:07:56,717 you know, many of you are going to take very different approaches to how you 134 00:07:56,800 --> 00:08:00,437 regulate AI, how you welcome it into your country, how you use it. 135 00:08:00,520 --> 00:08:03,237 But for sure, I think it is non-negotiable that you have to use it. 136 00:08:03,320 --> 00:08:07,797 The economic growth and benefit to people that can come from this, 137 00:08:07,880 --> 00:08:11,717 um the value to a country is too high to ignore. Uh there will be different ways 138 00:08:11,800 --> 00:08:15,877 people try to do this, but this is like, you know, the smartest people in the 139 00:08:15,960 --> 00:08:20,397 world uh decided to move to your country, 140 00:08:20,480 --> 00:08:23,917 to learn your language, um to work very hard on whatever problem 141 00:08:24,000 --> 00:08:28,477 you'd like, um to go discover new things that you need, and you can put as much 142 00:08:28,560 --> 00:08:31,637 of this to work as you want. Some of you will choose to build data 143 00:08:31,720 --> 00:08:34,397 centers, some of you will choose to rent data centers from other people because 144 00:08:34,480 --> 00:08:38,437 you don't want them in your country. I think all that's fine. Um but we want to 145 00:08:38,520 --> 00:08:42,677 help, we want to support, not just our company, our whole industry, like we 146 00:08:42,760 --> 00:08:48,197 want to bring this incredible magic of, 147 00:08:48,280 --> 00:08:51,397 you know, intelligence in a bottle, education in a bottle, 148 00:08:51,480 --> 00:08:54,317 um to all of you. I think this is like 149 00:08:54,400 --> 00:08:58,717 this is an incredible moment for the world. What people 150 00:08:58,800 --> 00:08:59,597 can 151 00:08:59,680 --> 00:09:03,677 do with this, and the economic 152 00:09:03,760 --> 00:09:08,677 boom that will come with it, the sort of individual increased quality of life 153 00:09:08,760 --> 00:09:13,517 people talk about, you know, care, and what that means, education what that 154 00:09:13,600 --> 00:09:17,517 means, services that only very rich people used to have access to, now 155 00:09:17,600 --> 00:09:21,997 anybody can have with AI. Um and the ability for people to just come up with 156 00:09:22,080 --> 00:09:24,197 ideas and get them to fruition so 157 00:09:24,280 --> 00:09:27,277 quickly, uh is going to be incredible. 158 00:09:27,360 --> 00:09:31,597 >> Do you Do you think that, um 159 00:09:31,680 --> 00:09:35,917 your company can create the context, 160 00:09:36,000 --> 00:09:39,877 um that people can use to create uh, 161 00:09:39,960 --> 00:09:41,597 their outcomes, or do you think there'll 162 00:09:41,680 --> 00:09:45,237 be an industry sort of surrounding you 163 00:09:45,320 --> 00:09:48,757 that will contextualize it? Because if you think about it, you say the the 164 00:09:48,840 --> 00:09:51,957 smartest people in the world move in to my neighborhood, and they're willing to 165 00:09:52,040 --> 00:09:55,997 work for me, but if I don't know what to ask them to do, that's not really a fun 166 00:09:56,080 --> 00:09:59,117 dinner party having the smartest people in the world sitting with you, and they 167 00:09:59,200 --> 00:10:01,037 can't really have a conversation. Right? 168 00:10:01,120 --> 00:10:03,677 So, how do we How do we think about context? 169 00:10:03,760 --> 00:10:06,717 >> That That's not going to come from us, and I don't think that should. Um, we 170 00:10:06,800 --> 00:10:09,877 are We will provide this this engine. 171 00:10:09,960 --> 00:10:12,117 You know, we will provide this 172 00:10:12,200 --> 00:10:14,997 incredible 173 00:10:15,080 --> 00:10:20,037 commodity of AI that can you can use for whatever you want, but the people of the 174 00:10:20,120 --> 00:10:23,077 world, and I have complete trust in the people of the world's creativity and 175 00:10:23,160 --> 00:10:25,637 ambition, um, will have to figure out 176 00:10:25,720 --> 00:10:28,437 what to ask it to do. What ideas to to 177 00:10:28,520 --> 00:10:32,997 to throw this at? What what problems and what context in a particular society or 178 00:10:33,080 --> 00:10:37,637 company is important to understand what to do. Um, so that won't be our role. I 179 00:10:37,720 --> 00:10:41,517 don't think we'd be good at it even if it were. Um, but we will be able to 180 00:10:41,600 --> 00:10:44,397 enable everyone around the world to bring that context and those questions 181 00:10:44,480 --> 00:10:49,797 and those ideas. >> So, should 182 00:10:49,880 --> 00:10:52,757 Should the governments of the world think of 183 00:10:52,840 --> 00:10:57,317 AI and companies, like your company, 184 00:10:57,400 --> 00:11:01,117 different than they thought about prior 185 00:11:01,200 --> 00:11:06,477 technology? Meaning, that if if what you're bringing them is, 186 00:11:06,560 --> 00:11:10,437 you know, the smartest people in the world, an educated 187 00:11:10,520 --> 00:11:15,197 capacity, then if they look at that as technology, and they 188 00:11:15,280 --> 00:11:17,917 uh, keep it out of their country, or they 189 00:11:18,000 --> 00:11:20,037 constrain it, are they Are they 190 00:11:20,120 --> 00:11:25,920 constraining their capacity to think and grow and build? 191 00:11:27,400 --> 00:11:31,237 >> I am always tempted to say this time is different. This is really special, you 192 00:11:31,320 --> 00:11:36,317 know, this is I and I suspect 193 00:11:36,400 --> 00:11:39,717 Well, to zoom out even further than that, as a kid 194 00:11:39,800 --> 00:11:44,677 my understanding of the world was that there were these few 195 00:11:44,760 --> 00:11:48,757 punctuated moments in history, each of which was a very different thing. You had 196 00:11:48,840 --> 00:11:52,357 um the agricultural revolution, and then you had the industrial revolution, and 197 00:11:52,440 --> 00:11:54,197 then you had the computer revolution. 198 00:11:54,280 --> 00:11:58,037 And these were all completely different, unrelated things, and there were not 199 00:11:58,120 --> 00:11:59,997 that many lessons to learn from them. 200 00:12:00,080 --> 00:12:03,557 And they were best understood in isolation. And even if you zoomed into 201 00:12:03,640 --> 00:12:06,157 one, pick the industrial revolution, um 202 00:12:06,240 --> 00:12:09,797 the right way to understand that was um 203 00:12:09,880 --> 00:12:13,637 20 or 30 important pieces of technology that were all developed over a period of 204 00:12:13,720 --> 00:12:16,597 a couple hundred years that came together in this important way. 205 00:12:16,680 --> 00:12:21,677 Uh as an adult, I look at it as you know, 206 00:12:21,760 --> 00:12:25,837 it has been a single revolution, a single exponential of technology. Each 207 00:12:25,920 --> 00:12:30,037 generation has built on the scaffolding that the generations before put a new 208 00:12:30,120 --> 00:12:33,557 brick in place. People have been able to see a or layer of scaffolding, to stick 209 00:12:33,640 --> 00:12:36,437 with the analogy, people have been able to climb and see a little higher and do 210 00:12:36,520 --> 00:12:40,397 the new thing. Um if you think about what it took to get 211 00:12:40,480 --> 00:12:44,717 us to this point where we can build AI, it was all of those previous revolutions 212 00:12:44,800 --> 00:12:47,757 stacked together, and it was the collective 213 00:12:47,840 --> 00:12:51,637 discovery of science throughout society and the global economy that we all 214 00:12:51,720 --> 00:12:54,637 contribute to that made this incredible 215 00:12:54,720 --> 00:12:59,757 thing possible and now happen. And it is tempting whenever you're in the 216 00:12:59,840 --> 00:13:02,237 moment to say, "This is the last revolution. This is the biggest 217 00:13:02,320 --> 00:13:05,357 revolution." There will be more after this. There will be 218 00:13:05,440 --> 00:13:08,037 bigger things after this. People will use AI to invent whatever thing that 219 00:13:08,120 --> 00:13:12,117 comes next. It's probably very difficult for us to imagine right now. 220 00:13:12,200 --> 00:13:17,157 Um and and I think the right answer is to look 221 00:13:17,240 --> 00:13:21,557 at the history of technology as one continuous exponential. The history of 222 00:13:21,640 --> 00:13:25,317 the economy and society as one continuous exponential. Um 223 00:13:25,400 --> 00:13:29,397 and I really mean that. It's not just the technology, it's this social 224 00:13:29,480 --> 00:13:32,877 fabric and the kind of complex structure of the world that allows this technology 225 00:13:32,960 --> 00:13:35,437 to flourish and people to figure out what to do with it. 226 00:13:35,520 --> 00:13:38,397 And you have to embrace it continuously all 227 00:13:38,480 --> 00:13:41,677 the way through. If you miss out on any part of it, um you're not ready for the 228 00:13:41,760 --> 00:13:46,677 next thing that will come. If you If you don't enable your people and your 229 00:13:46,760 --> 00:13:50,917 businesses with AI, uh they will not invent the next level 230 00:13:51,000 --> 00:13:54,637 of the scaffolding, but there will be more things. And And I think it is 231 00:13:54,720 --> 00:13:58,317 tempting to say this one is different and it clearly is really important. Um 232 00:13:58,400 --> 00:14:01,197 but I would view it as a continuum. 233 00:14:01,280 --> 00:14:02,957 So, 234 00:14:03,040 --> 00:14:06,117 >> we all read about the the various 235 00:14:06,200 --> 00:14:11,197 companies at the frontier. Can you Can you sort of differentiate them for us? 236 00:14:11,280 --> 00:14:12,837 What What do you 237 00:14:12,920 --> 00:14:15,637 What do you view OpenAI as best at? What 238 00:14:15,720 --> 00:14:20,517 What's the, you know, your DNA, if you will, that separates you from the 239 00:14:20,600 --> 00:14:22,397 others? And And what are you great at? 240 00:14:22,480 --> 00:14:26,877 And just give us that sort of from your perspective. 241 00:14:26,960 --> 00:14:29,280 >> Um 242 00:14:30,120 --> 00:14:33,397 I think from a from what matters to this room, um we 243 00:14:33,480 --> 00:14:35,597 are the pragmatist's interests. 244 00:14:35,680 --> 00:14:41,397 We We realize like we want to be We want to be accurate in how we talk 245 00:14:41,480 --> 00:14:43,557 about what we see in the future. We don't want to fall into the trap of 246 00:14:43,640 --> 00:14:46,917 blind optimism. We don't want to fall into the trap of doomerism. This is 247 00:14:47,000 --> 00:14:50,677 clearly complex. There's clearly a combination of both excitement and 248 00:14:50,760 --> 00:14:54,877 anxiety, of transformation and risks that the world has to 249 00:14:54,960 --> 00:14:58,437 navigate. Um but 250 00:14:58,520 --> 00:15:02,477 by getting pulled into either extreme, we are not going to successfully figure 251 00:15:02,560 --> 00:15:04,397 out the challenges in front of us. 252 00:15:04,480 --> 00:15:07,957 Um we want to be a reliable partner. Uh 253 00:15:08,040 --> 00:15:11,877 we want Like we really do view what we are doing as a platform, as a service to 254 00:15:11,960 --> 00:15:14,957 the world. We realize that different people around the world are going to use 255 00:15:15,040 --> 00:15:18,157 it very differently, and we want to 256 00:15:18,240 --> 00:15:23,197 put sufficient constraints on our system to avoid catastrophic risks, but we 257 00:15:23,280 --> 00:15:26,357 really do want to respect all of your sovereignty and that you're going to use 258 00:15:26,440 --> 00:15:30,677 it in very different ways and we personally may not always like it, um 259 00:15:30,760 --> 00:15:34,877 but I don't like everything that people do with electricity either and I 260 00:15:34,960 --> 00:15:36,677 still think everybody should get to have it. 261 00:15:36,760 --> 00:15:40,080 Um so 262 00:15:41,000 --> 00:15:45,237 Oh, and then also I would say like we want the whole global economy to 263 00:15:45,320 --> 00:15:49,117 succeed. We do not want to be the only company. Um we do not want to suck up 264 00:15:49,200 --> 00:15:52,677 all the value. We do not think that is good stability nor do we think that's 265 00:15:52,760 --> 00:15:56,237 possible for the way that people need to flourish and people need to stay at the 266 00:15:56,320 --> 00:16:00,757 center of this. I'll walk into that one more thing after this. Um 267 00:16:00,840 --> 00:16:05,157 and we want to like enable every entrepreneur, every government, every 268 00:16:05,240 --> 00:16:07,917 big company in the world to thrive with this technology. 269 00:16:08,000 --> 00:16:11,437 Uh the last thing I'll say that I think differentiates us from some of the other 270 00:16:11,520 --> 00:16:16,117 companies is we are very firmly and very proudly on team humanity. Um we would 271 00:16:16,200 --> 00:16:18,357 like to build tools for people. We would 272 00:16:18,440 --> 00:16:21,477 like people's lives to be better. Um we 273 00:16:21,560 --> 00:16:23,637 think that people need to be at the 274 00:16:23,720 --> 00:16:27,197 center of the economy and society and 275 00:16:27,280 --> 00:16:30,957 global decision-making um and that people should run the future 276 00:16:31,040 --> 00:16:34,277 not AI. So 277 00:16:34,360 --> 00:16:37,797 >> it seems that entrepreneurship 278 00:16:37,880 --> 00:16:42,837 should be really the fire starter of all this, right? You're Are you the 279 00:16:42,920 --> 00:16:44,797 you know, the the kindling of it? I 280 00:16:44,880 --> 00:16:46,437 mean, maybe sort of tell me how 281 00:16:46,520 --> 00:16:50,237 entrepreneurship and and uh 282 00:16:50,320 --> 00:16:52,997 uh Open the Eyes sort of work together. 283 00:16:53,080 --> 00:16:56,637 >> B- Before OpenAI, I used to run a startup accelerator. 284 00:16:56,720 --> 00:17:00,997 And I had always been interested not always after I 285 00:17:01,080 --> 00:17:03,997 came to college, I'd always been interested in entrepreneurship and then 286 00:17:04,080 --> 00:17:05,597 having had this incredible front row 287 00:17:05,680 --> 00:17:09,597 seat, I really became convinced that 288 00:17:09,680 --> 00:17:12,077 entrepreneurship was 289 00:17:12,160 --> 00:17:14,517 maybe the most important 290 00:17:14,600 --> 00:17:17,557 novel force in all of capitalism. Um 291 00:17:17,640 --> 00:17:21,237 and I fell in love with it. And 292 00:17:21,320 --> 00:17:22,917 watching what 293 00:17:23,000 --> 00:17:25,437 new companies were able to do 294 00:17:25,520 --> 00:17:29,997 and watching how large of a percentage of the new ideas 295 00:17:30,080 --> 00:17:34,157 that mattered in the tech industry over the last, you know, 20 years or 296 00:17:34,240 --> 00:17:37,277 whatever came from startups and not big companies. 297 00:17:37,360 --> 00:17:39,237 Um now having seen this in other 298 00:17:39,320 --> 00:17:45,880 industries, I am a huge believer in entrepreneurship. I also think that 299 00:17:46,720 --> 00:17:49,517 you know, like I talk to a lot of people around the world and they would much 300 00:17:49,600 --> 00:17:53,157 rather run a great lifestyle business than go work as like a corporate cog. 301 00:17:53,240 --> 00:17:56,997 And I think that's great, too. Like I think that's just we should enable that. 302 00:17:57,080 --> 00:18:01,437 Um one of the things I got excited about early with AI, of course AI is going to 303 00:18:01,520 --> 00:18:04,997 transform big companies. And of course AI is going to transform governments. 304 00:18:05,080 --> 00:18:08,400 But what AI can do 305 00:18:08,560 --> 00:18:11,317 for entrepreneurs and for people that 306 00:18:11,400 --> 00:18:13,877 want to take an idea and bring it into 307 00:18:13,960 --> 00:18:17,037 the world quickly, um is really quite 308 00:18:17,120 --> 00:18:21,117 amazing. When I was running the startup accelerator, it lasted for 3 months. And 309 00:18:21,200 --> 00:18:23,037 the work >> It lasted for 3 months. 310 00:18:23,120 --> 00:18:27,597 >> You ran it for 3 months. And the work that we expected a startup to accomplish 311 00:18:27,680 --> 00:18:32,157 in 3 months is now probably doable in like 17 minutes 312 00:18:32,240 --> 00:18:35,517 with Codex. This is just amazing. The world has not 313 00:18:35,600 --> 00:18:39,517 caught up with what this is going to mean, not just for the economic value 314 00:18:39,600 --> 00:18:44,477 that gets created from that, but how quickly people can test ideas, build 315 00:18:44,560 --> 00:18:46,637 things, get feedback from your customers. We're all just going to get 316 00:18:46,720 --> 00:18:49,357 much better stuff. 317 00:18:49,440 --> 00:18:51,637 >> So you used the term 318 00:18:51,720 --> 00:18:57,037 I want to sort of flush it out a bit a bit, non-negotiable. 319 00:18:57,120 --> 00:19:00,797 Can you maybe uh help everybody understand what you 320 00:19:00,880 --> 00:19:03,077 mean about that and and what that means 321 00:19:03,160 --> 00:19:08,680 for each of our partner ministers who are here today with us? 322 00:19:10,480 --> 00:19:13,117 >> Uh again, I'm biased. I'm trying not not my 323 00:19:13,200 --> 00:19:15,797 book here. Forgive me for this analogy. 324 00:19:15,880 --> 00:19:19,557 Um, I think it would be approximately as bad 325 00:19:19,640 --> 00:19:23,757 of an idea to say we're not going to have AI in our country as it was to say 326 00:19:23,840 --> 00:19:26,957 we're not going to have electricity in our country, you know, back 100 plus 327 00:19:27,040 --> 00:19:31,077 years ago. I think this is something at that level. Um, this is something that 328 00:19:31,160 --> 00:19:33,837 every person and every business and 329 00:19:33,920 --> 00:19:35,837 really the entire 330 00:19:35,920 --> 00:19:39,357 machine of society is going to need to 331 00:19:39,440 --> 00:19:42,237 deliver the quality of life 332 00:19:42,320 --> 00:19:47,277 that your citizens should demand. Now, electricity also came with some 333 00:19:47,360 --> 00:19:50,637 complexities and it took a little while for us to figure out how to make it 334 00:19:50,720 --> 00:19:54,477 safe, but we made fairly rapid progress as a society to make it generally pretty 335 00:19:54,560 --> 00:19:59,157 safe even though it can do a lot of bad things. Um, and at this point, we barely 336 00:19:59,240 --> 00:20:03,437 think about electricity. You know, we turn these lights on, we expect them to work. 337 00:20:03,520 --> 00:20:06,717 You know, if this the power shut off and I had to scream cuz this microphone was 338 00:20:06,800 --> 00:20:10,397 not working, like I would be annoyed about that. You 339 00:20:10,480 --> 00:20:12,757 I won't go through all the examples. I see all the phones on the table. Like it 340 00:20:12,840 --> 00:20:15,677 just It would be hard for society to function well 341 00:20:15,760 --> 00:20:18,717 without electricity at this point. But we don't think about it much. We think 342 00:20:18,800 --> 00:20:20,597 about, you know, there's light at night. 343 00:20:20,680 --> 00:20:23,517 That's great. We can hear these people speak in the 344 00:20:23,600 --> 00:20:25,877 room without them having to yell. That's nice. Our phones can do these amazing 345 00:20:25,960 --> 00:20:27,517 things for us. We care about our phones. 346 00:20:27,600 --> 00:20:30,317 We don't think too much about the electricity that gets generated to power 347 00:20:30,400 --> 00:20:33,077 them. Um, and I think the same thing is going to 348 00:20:33,160 --> 00:20:37,437 happen with AI. We won't My prediction is we won't be talking 349 00:20:37,520 --> 00:20:41,517 about AI as much as we do in a decade from now because we'll just 350 00:20:41,600 --> 00:20:46,437 expect it everywhere. You know, like a kid growing up today will never be 351 00:20:46,520 --> 00:20:51,917 smarter than AI, but he or she will also never 352 00:20:52,000 --> 00:20:56,197 have understood a world where every product and service 353 00:20:56,280 --> 00:20:58,877 that they interact with is not really smart and really capable and really 354 00:20:58,960 --> 00:21:02,317 helpful. And because we'll just be so used to that, 355 00:21:02,400 --> 00:21:05,197 we won't think about the AI. We'll just be like, "Yeah, I want my products to be 356 00:21:05,280 --> 00:21:08,757 good." You know, if I have an idea, I want to be able to like make it right 357 00:21:08,840 --> 00:21:11,877 away. I want this software to be created instantly for me. 358 00:21:11,960 --> 00:21:16,157 Um if I have a disease, like I expect a cure to get discovered very quickly. 359 00:21:16,240 --> 00:21:20,757 Uh so I think countries need to need to embrace this. Again, there's many ways 360 00:21:20,840 --> 00:21:23,957 to do it. People will set different rules. People will decide on 361 00:21:24,040 --> 00:21:28,277 what they want to build versus rent, but I think everyone will need the 362 00:21:28,360 --> 00:21:33,357 this new commodity. >> If um if we look forward 363 00:21:33,440 --> 00:21:36,717 5 years and and we come back to North Carolina 364 00:21:36,800 --> 00:21:39,717 all together um 365 00:21:39,800 --> 00:21:44,800 what what will be holding us back, do you think, in your opinion? 366 00:21:46,920 --> 00:21:50,317 >> Well, there are clearly challenges that 367 00:21:50,400 --> 00:21:54,157 we have to get right along the way. There will be specifics. 368 00:21:54,240 --> 00:21:58,437 For example, I'm sure cybersecurity is a top-of-mind issue for many of you. 369 00:21:58,520 --> 00:22:01,637 Uh we work with many of the governments in the room. Uh 370 00:22:01,720 --> 00:22:06,037 we work with the state of North Carolina. Uh we work with companies big and small. 371 00:22:06,120 --> 00:22:10,917 And we are I think we have a big challenge in front 372 00:22:11,000 --> 00:22:14,917 of us. Like to not to not try to sidestep this one. I think 373 00:22:15,000 --> 00:22:18,557 some things are going to go very wrong with cybersecurity unless people act 374 00:22:18,640 --> 00:22:21,557 quite urgently. There will be other challenges in the 375 00:22:21,640 --> 00:22:24,797 next 5 years. People talk about um biosecurity and the the things we're 376 00:22:24,880 --> 00:22:28,597 going to face there. Um 377 00:22:28,680 --> 00:22:30,637 there will be bigger ones yet to come. 378 00:22:30,720 --> 00:22:35,677 And I think if we don't successfully navigate those um 379 00:22:35,760 --> 00:22:37,557 society will not be able to adopt AI to 380 00:22:37,640 --> 00:22:41,197 the degree or to the speed with which 381 00:22:41,280 --> 00:22:46,197 people deserve. So, getting successfully navigating the risks and 382 00:22:46,280 --> 00:22:49,597 downsides in front of us is something that could go very wrong, 383 00:22:49,680 --> 00:22:52,997 and I think we'll set this technology back. Could set this technology back a 384 00:22:53,080 --> 00:22:56,757 great deal. Um another thing that could go wrong is too 385 00:22:56,840 --> 00:23:01,200 much concentration of power. Um 386 00:23:01,800 --> 00:23:05,677 AI should be in many ways AI should be an equalizing 387 00:23:05,760 --> 00:23:08,797 force in the world. You talked about what it means to have like a reset on 388 00:23:08,880 --> 00:23:11,717 the educational capacity available to everybody. 389 00:23:11,800 --> 00:23:16,117 Um you can also see worlds where AI 390 00:23:16,200 --> 00:23:19,597 concentrates because some people make early investments and 391 00:23:19,680 --> 00:23:23,957 others don't and there's, you know, bad inequities with computer whatever. Um 392 00:23:24,040 --> 00:23:27,677 that'd be bad. So I think if we fail to make AI abundant, it 393 00:23:27,760 --> 00:23:29,637 ends up being something that is very 394 00:23:29,720 --> 00:23:31,917 highly competed over, rich people spend 395 00:23:32,000 --> 00:23:37,280 a huge amount of money on, and um 396 00:23:37,600 --> 00:23:41,437 that'd be a bad outcome. So I think we need to make it abundant so that 397 00:23:41,520 --> 00:23:44,557 we get the the benefits of lifting everybody up and 398 00:23:44,640 --> 00:23:46,437 not much worse than the inequality. 399 00:23:46,520 --> 00:23:50,477 >> Right. As as you said, if if the smartest people in the world 400 00:23:50,560 --> 00:23:54,477 learned your language and all agreed to move to your country 401 00:23:54,560 --> 00:23:58,957 and work on whatever you wanted them to work on, that that has to be the most 402 00:23:59,040 --> 00:24:03,957 incredible thing. Do you think that 403 00:24:04,040 --> 00:24:07,477 since you'll have all the smartest people speaking all those languages, 404 00:24:07,560 --> 00:24:10,757 can't they figure out a way, do you think over time, 405 00:24:10,840 --> 00:24:12,957 to deliver the same amount of compute 406 00:24:13,040 --> 00:24:16,397 more inexpensively without the same use 407 00:24:16,480 --> 00:24:18,597 of power, without the same I mean, 408 00:24:18,680 --> 00:24:21,237 aren't these the kind of things that are 409 00:24:21,320 --> 00:24:26,237 going to happen that will democratize it by sort of by itself in a way? 410 00:24:26,320 --> 00:24:29,437 >> It it will still require a huge investment, 411 00:24:29,520 --> 00:24:33,797 but yes, we will continue to make things vastly more efficient. And we'll figure 412 00:24:33,880 --> 00:24:36,597 out to make more and better chips, and we'll figure out 413 00:24:36,680 --> 00:24:40,757 to make smarter algorithms. All of that will happen. But to put some numbers on 414 00:24:40,840 --> 00:24:43,597 the magnitude of how much we we need this to happen, 415 00:24:43,680 --> 00:24:45,517 again, I don't have these numbers in front of me. I'm not going to get this 416 00:24:45,600 --> 00:24:49,837 exactly right. Um but forgive me for the approximation. 417 00:24:49,920 --> 00:24:53,997 At the beginning of 2020, um the token leader in the world was an 418 00:24:54,080 --> 00:24:58,717 OpenAI employee using roughly 100,000 tokens a month. 419 00:24:58,800 --> 00:25:03,677 Um in the middle of 2026, the token leader that I know of, there 420 00:25:03,760 --> 00:25:07,037 may be somebody using more than this at outside of 421 00:25:07,120 --> 00:25:09,437 OpenAI, um was using hundreds of 422 00:25:09,520 --> 00:25:11,637 billions of tokens a month. 423 00:25:11,720 --> 00:25:14,597 So, you know, a multi-million X increase 424 00:25:14,680 --> 00:25:18,677 in 6 and 1/2 years, and the average person on Earth had gone from using 425 00:25:18,760 --> 00:25:21,917 effectively zero tokens to 100,000 tokens a month. That was the global per 426 00:25:22,000 --> 00:25:25,797 capita thing. If we just naively project this forward 427 00:25:25,880 --> 00:25:30,157 again, which seems like a reasonable thing to do to me, um and I hate tokens 428 00:25:30,240 --> 00:25:34,077 as a unit here. It's a dumb one. No one should care. But, like, let's just take 429 00:25:34,160 --> 00:25:35,957 that as shorthand for amount of 430 00:25:36,040 --> 00:25:39,277 intelligence, then 431 00:25:39,360 --> 00:25:44,157 in another 6 and 1/2 years, the average per capita use of tokens will go up to 432 00:25:44,240 --> 00:25:46,757 Let's round it down to 100 billion tokens a month. 433 00:25:46,840 --> 00:25:52,320 And the person using the most in the world will be using, 434 00:25:52,520 --> 00:25:56,037 you know, a 100 quadrillion tokens a month. Uh 435 00:25:56,120 --> 00:25:59,317 or uh 1 quadrillion 100 trillion. Like >> Mhm. 436 00:25:59,400 --> 00:26:02,637 >> No, I was right. 100 trillion tokens a month. 100 quadrillion tokens a month. 437 00:26:02,720 --> 00:26:07,917 Um even with great gains on efficiency, 438 00:26:08,000 --> 00:26:11,197 we are going to need to build much more infrastructure 439 00:26:11,280 --> 00:26:15,877 unless we want this to be a highly priced commodity, which we don't. We 440 00:26:15,960 --> 00:26:17,997 want this to be abundant. We want this to be inexpensive. We want people to 441 00:26:18,080 --> 00:26:20,597 just be able to use this for whatever they need. 442 00:26:20,680 --> 00:26:23,557 So, that's the driving force. If we 443 00:26:23,640 --> 00:26:27,237 if we globally can build enough, 444 00:26:27,320 --> 00:26:29,877 then it will intelligence will be 445 00:26:29,960 --> 00:26:31,797 broadly distributed, 446 00:26:31,880 --> 00:26:35,037 uh more democratized across the world. 447 00:26:35,120 --> 00:26:38,597 Uh the world will be educated. Well, 448 00:26:38,680 --> 00:26:41,157 they'll have the access to education. 449 00:26:41,240 --> 00:26:42,797 >> Is that sort of the >> Absolutely. 450 00:26:42,880 --> 00:26:44,837 >> To To go back to the analogy of 451 00:26:44,920 --> 00:26:47,677 electricity, um there's this amazing 452 00:26:47,760 --> 00:26:51,197 chart of what it cost to have an hour of 453 00:26:51,280 --> 00:26:55,237 light at night. And 454 00:26:55,320 --> 00:26:58,917 there was a time, one of the earliest data points you can find, is that it it 455 00:26:59,000 --> 00:27:01,837 took an average of 5 hours of labor at 456 00:27:01,920 --> 00:27:03,477 the wage of the time to be able to pay 457 00:27:03,560 --> 00:27:06,837 for 1 hour of light at night. 458 00:27:06,920 --> 00:27:10,837 Electricity came along, it fell, we figured out how to make more 459 00:27:10,920 --> 00:27:13,917 electricity, so there was like, you know, the supply and demand change 460 00:27:14,000 --> 00:27:18,997 curve, it fell again, we figured out how to make vastly more efficient light 461 00:27:19,080 --> 00:27:22,717 bulbs, we went from incandescent lights to LEDs, a whole bunch of other things 462 00:27:22,800 --> 00:27:26,837 happened. And now an hour of light at night is not 463 00:27:26,920 --> 00:27:29,997 something that people think of at all about the cost. I don't even know 464 00:27:30,080 --> 00:27:33,437 what the number is, but someone here can do it. It's it's like very cheap to run, 465 00:27:33,520 --> 00:27:35,717 you know, a 5-W bulb for an hour. 466 00:27:35,800 --> 00:27:41,077 Um the If that hadn't happened, if we and all 467 00:27:41,160 --> 00:27:44,157 of those things together had to happen, lots of electricity generation, huge 468 00:27:44,240 --> 00:27:46,597 innovation on the efficiency, all of these things. 469 00:27:46,680 --> 00:27:50,877 If that hadn't happened, then rich people get light at night, and they can 470 00:27:50,960 --> 00:27:53,397 study more, they have a compounding effect, they can learn more, whatever, 471 00:27:53,480 --> 00:27:57,477 and and, you know, poor people don't. So, I think it's been 472 00:27:57,560 --> 00:28:00,877 very good that energy has become abundant and that light at night has 473 00:28:00,960 --> 00:28:04,597 become relatively, like, shockingly inexpensive. 474 00:28:04,680 --> 00:28:08,117 Um If that same thing happens with AI 475 00:28:08,200 --> 00:28:12,717 and anybody on Earth, rich, poor, whatever, can use AI to get 476 00:28:12,800 --> 00:28:16,117 the abundant services that they need and to express their ideas and to start 477 00:28:16,200 --> 00:28:17,957 their businesses and whatever else, and 478 00:28:18,040 --> 00:28:23,277 we fulfill our mission of making AI so inexpensive and so smart that people are 479 00:28:23,360 --> 00:28:25,677 just like they think about it like they think of 480 00:28:25,760 --> 00:28:28,757 the cost of an hour at night of light at night, I think that'd be really 481 00:28:28,840 --> 00:28:32,317 wonderful. And then this huge reset and democratization of education you're 482 00:28:32,400 --> 00:28:35,917 talking about can happen. >> So, 483 00:28:36,000 --> 00:28:38,237 are you, you know, there are 484 00:28:38,320 --> 00:28:43,517 Look, there are doomers out there who worry about everything, but we know that 485 00:28:43,600 --> 00:28:47,397 if you look back through the industrial revolution, as you said, it it always 486 00:28:47,480 --> 00:28:49,837 seems to work out fine. Are you 487 00:28:49,920 --> 00:28:52,957 uh fundamentally optimistic to what your 488 00:28:53,040 --> 00:28:57,637 products will do to the world and will serve the world? 489 00:28:57,720 --> 00:29:02,357 >> I I think part of the reason it tends to work out fine is people 490 00:29:02,440 --> 00:29:05,757 worry and they stress and they look at the downsides and they try to address 491 00:29:05,840 --> 00:29:10,317 those and they try to learn how we avoid accidents. So, if we go back to an early 492 00:29:10,400 --> 00:29:13,477 example even before electricity of a technology 493 00:29:13,560 --> 00:29:18,797 of fire was an incredible incredible for human evolution, incredible for letting 494 00:29:18,880 --> 00:29:21,837 us cook our food and heat our homes at night and whatever else. It also burnt 495 00:29:21,920 --> 00:29:24,517 down whole cities a few times. 496 00:29:24,600 --> 00:29:28,357 And people really worried about that cuz we don't want cities to burn down. 497 00:29:28,440 --> 00:29:30,757 And we figured out fire safety. 498 00:29:30,840 --> 00:29:35,677 And through a lot of human ingenuity, um through a lot of 499 00:29:35,760 --> 00:29:38,517 good policy definitely through learning from some 500 00:29:38,600 --> 00:29:43,197 very bad accidents, we figured out how to make it safer. Now, had we just said 501 00:29:43,280 --> 00:29:46,237 like, "Yeah, you know, some cities are going to burn down. Whatever, we got 502 00:29:46,320 --> 00:29:49,157 other problems to work on." Probably wouldn't have figured that out and that 503 00:29:49,240 --> 00:29:52,757 wouldn't have been as good. Um so, 504 00:29:52,840 --> 00:29:57,357 I think worry is warranted. Worry is how we anticipate problems and 505 00:29:57,440 --> 00:30:00,597 how we from a matter of policy and a matter of 506 00:30:00,680 --> 00:30:04,277 technology figure out how to address them. But I am optimistic that we can 507 00:30:04,360 --> 00:30:08,277 mitigate the downsides of this technology. And if you look at what's happened, 508 00:30:08,360 --> 00:30:10,397 you know, if you look at what the doomers were 509 00:30:10,480 --> 00:30:12,837 saying 5 years ago about the chances of 510 00:30:12,920 --> 00:30:17,477 us being able to make AI of this level, of this sort of near human, sometimes 511 00:30:17,560 --> 00:30:21,437 superhuman capacity level, you know, broadly considered 512 00:30:21,520 --> 00:30:24,237 safe and reliable. Not perfect. There have been accidents, there will be 513 00:30:24,320 --> 00:30:27,157 accidents, of course. I don't think most people would have 514 00:30:27,240 --> 00:30:28,997 thought it was going to go this well 515 00:30:29,080 --> 00:30:32,437 as it has so far. And I I believe in 516 00:30:32,520 --> 00:30:37,397 human ingenuity and the ability for to solve very difficult, very complex 517 00:30:37,480 --> 00:30:40,597 problems and I think I think we'll keep doing that. 518 00:30:40,680 --> 00:30:42,837 And I think the economic 519 00:30:42,920 --> 00:30:46,237 value and the personal value is so huge. 520 00:30:46,320 --> 00:30:49,517 Um I met with someone just this morning and told and again I'm sure you've all 521 00:30:49,600 --> 00:30:54,197 heard stories like this that they wouldn't be alive without ChatGPT um and 522 00:30:54,280 --> 00:30:57,837 that they had a a very difficult disease and you know got a diagnosis the doctors 523 00:30:57,920 --> 00:31:00,837 weren't able to give them. We shouldn't take that away. 524 00:31:00,920 --> 00:31:02,757 >> No. No. >> Like we need to figure out how to make 525 00:31:02,840 --> 00:31:05,837 the technology safe so that people can have more of that. 526 00:31:05,920 --> 00:31:08,597 >> Great. Sam, thank you for joining us. It was 527 00:31:08,680 --> 00:31:12,117 really an incredible conversation. Um 528 00:31:12,200 --> 00:31:16,717 2:00 so 2:00 p.m. in the ministerial room 529 00:31:16,800 --> 00:31:22,677 where we will have uh Alex Karp will be our next guest uh speaking to us. 530 00:31:22,760 --> 00:31:27,637 >> He's a good one. >> He's a good one. But uh thank you Sam for joining us. We We 531 00:31:27,720 --> 00:31:29,617 really appreciate it. Thank you. 532 00:31:29,700 --> 00:31:33,360 >> [applause] >> Thanks Sam.